Convolutional Layer Implementation To Classify Malware in Banking Financial Services Industry

UNIVERSITY OF INDONESIA
CONVOLUTIONAL LAYER IMPLEMENTATION TO CLASSIFY MALWARE IN

BANKING FINANCIAL SERVICES INDUSTRY
THESIS
AZIS S PRASETYOTOMO
11060224390
FACULTY OF COMPUTER SCIENCE

MAGISTER INFORMATION TECHNOLOGY
JAKARTA
JULI 2017
Abstract
According to AV vendors malicious software has been growing exponentially

last years. One of the main reasons for these high volumes is that in order
to evade detection, malware authors started using polymorphic and meta-
morphic techniques. As a result, traditional signature-based approaches to
detect malware are being insufficient against new malware and the catego-
rization of malware samples had become essential to know the basis of the
behavior of malware and to fight back cybercriminals.
During the last decade, solutions that fight against malicious software had
begun using machine learning approaches. Unfortunately, there are few open-
source datasets available for the academic community. One of the biggest
datasets available was released last year in a competition hosted on Kag-
gle with data provided by Microsoft for the Big Data Innovators Gathering
(BIG 2015). This thesis presents two novel and scalable approaches using
Convolutional Neural Networks (CNNs) to assign malware to its correspond-
ing family. On one hand, the first approach makes use of CNNs to learn a
feature hierarchy to discriminate among samples of malware represented as
gray-scale images. On the other hand, the second approach uses the CNN
architecture introduced by Yoon Kim [12] to classify malware samples accord-
ing their x86 instructions. The proposed methods achieved an improvement
of 93.86% and 98,56% with respect to the equal probability benchmark.
Acknowledgments
I would first like to thank my family, especially Mom, for the continuous
support she has given me throughout my time in graduate school. Second,
I would like to express my gratitude to my supervisor, Dr. Javier Béjar for
their guidance during the course of this thesis.
1
Contents
1 Introduction 8
1.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2 Background 14
2.1 Artificial Neural Networks . . . . . . . . . . . . . . . . . . . . 14
2.1.1 Perceptrons . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1.2 Sigmoid neuron . . . . . . . . . . . . . . . . . . . . . . 16
2.1.3 Loss function . . . . . . . . . . . . . . . . . . . . . . . 16
2.1.4 Gradient Descent Algorithm . . . . . . . . . . . . . . . 17
2.1.5 Backpropagation . . . . . . . . . . . . . . . . . . . . . 19
2.2 Convolutional Neural Networks . . . . . . . . . . . . . . . . . 21
2.2.1 Local connectivity . . . . . . . . . . . . . . . . . . . . 22
2.2.2 Convolutional Layer . . . . . . . . . . . . . . . . . . . 22
2.2.3 Pooling Layer . . . . . . . . . . . . . . . . . . . . . . . 23
2.3 Overfitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.1 Regularization . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.2 Dropout . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3.3 Artificially expanding the training data . . . . . . . . . 26
2.4 Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.1 ReLU units . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.2 Gradient Descent Optimization Algorithms . . . . . . . 29
2
CONTENTS
3 State of the Art 33
4 Microsoft Malware Classification Challenge 39

4.1 What’s Kaggle? . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 Microsoft Malware Classification Challenge . . . . . . . . . . . 40
4.2.1 Bytes file . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.2.2 ASM file . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3 Winner’s solution . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.4 Novel Feature Extraction, Selection and Fusion for Effective
Malware Family Classification . . . . . . . . . . . . . . . . . . 47
4.5 Deep Learning Frameworks . . . . . . . . . . . . . . . . . . . . 51
5 Learning Feature Extractors from Malware Images 53

5.1 Visualizing malware as gray-scale images . . . . . . . . . . . . 54
5.1.1 Malware families . . . . . . . . . . . . . . . . . . . . . 55
5.2 CNN Architectures . . . . . . . . . . . . . . . . . . . . . . . . 59
5.2.1 CNN A: 1C 1D . . . . . . . . . . . . . . . . . . . . . . 61
5.2.2 CNN B: 2C 1D . . . . . . . . . . . . . . . . . . . . . . 62
5.2.3 CNN C: 3C 2D . . . . . . . . . . . . . . . . . . . . . . 64
5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3.1 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3.2 Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
6 Convolutional Neural Networks for Classification of Malware

Disassembly Files 72
6.1 Representing Opcodes as Word Embeddings . . . . . . . . . . 74
6.1.1 Skip-Gram model . . . . . . . . . . . . . . . . . . . . . 75
6.2 Convolutional Neural Network Architecture . . . . . . . . . . 79
6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.3.1 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.3.2 Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3
CONTENTS
7 Conclusions 90
7.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4
List of Figures
2.1 Effects of different learning rates . . . . . . . . . . . . . . . . . 18

2.2 AlexNet architecture . . . . . . . . . . . . . . . . . . . . . . . 21
2.3 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Max pooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.5 ReL and sigmoid functions comparison . . . . . . . . . . . . . 29
3.1 Most frequent 14 opcodes for goodware . . . . . . . . . . . . . 34

3.2 Most frequent 14 opcodes for malware . . . . . . . . . . . . . 34
3.3 Outline of Invencea’s Malware Detection Framework . . . . . . 36
3.4 Visualizing Malware as an Image . . . . . . . . . . . . . . . . 38
4.1 Malware Classification Challenge: dataset . . . . . . . . . . . 40

4.2 Snapshot of one bytes file . . . . . . . . . . . . . . . . . . . . 41
4.3 Snapshot of one assembly code file . . . . . . . . . . . . . . . 42
4.4 Top 10 opcodes in the training dataset . . . . . . . . . . . . . 44
4.5 Average of opcodes per malware family . . . . . . . . . . . . . 44
5.1 Visualizing Malware as a Gray-Scale Image . . . . . . . . . . . 54

5.2 Rammit samples . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.3 Lollipop samples . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.4 Kelihos_ver3 samples . . . . . . . . . . . . . . . . . . . . . . . 56
5.5 Vundo samples . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.6 Simda samples . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5
LIST OF FIGURES
5.7 Tracur samples . . . . . . . . . . . . . . . . . . . . . . . . . . 57

5.8 Kelihos_ver1 samples . . . . . . . . . . . . . . . . . . . . . . . 57
5.9 Obfuscator.DCY samples . . . . . . . . . . . . . . . . . . . . . 57
5.10 Gatak samples . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.11 Shallow Approach . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.12 Overview architecture A: 1C 1D . . . . . . . . . . . . . . . . . 62
5.13 Overview architecture B: 2C 1D . . . . . . . . . . . . . . . . . 64
5.14 Overview architecture C: 3C 2D . . . . . . . . . . . . . . . . . 66
5.15 Approach A: CNNs training results . . . . . . . . . . . . . . . 68
6.1 Yoon Kim model architecture . . . . . . . . . . . . . . . . . . 73

6.2 Skip-gram model architecture . . . . . . . . . . . . . . . . . . 75
6.3 t-SNE representation of the word embeddings . . . . . . . . . 78
6.4 CNN Embedding layer output . . . . . . . . . . . . . . . . . . 80
6.5 CNN Convolutional layer output . . . . . . . . . . . . . . . . . 81
6.6 CNN Max-pooling & output layer . . . . . . . . . . . . . . . . 82
6.7 Heuristic Search: Learning Rate . . . . . . . . . . . . . . . . . 84
6.8 Heuristic Search: Embedding size . . . . . . . . . . . . . . . . 84
6.9 Heuristic Search: #Filters . . . . . . . . . . . . . . . . . . . . 85
6.10 Heuristic Search: Filter Sizes . . . . . . . . . . . . . . . . . . . 86
6.11 Approach B: CNNs training results . . . . . . . . . . . . . . . 87
6
List of Tables
4.1 Number of samples per class with 0 instructions . . . . . . . . 45

4.2 Winner’s solution: confusion matrix . . . . . . . . . . . . . . . 47
4.3 List of feature categories and their evaluation in XGBoost . . 49
5.1 CNN 1C 1D: confusion matrix . . . . . . . . . . . . . . . . . . 68

6.1 CNN without pretrained word embeddings: confusion matrix . 87

6.2 CNN with pretrained word embeddings: confusion matrix . . . 88
6.3 Approach B: Test scores . . . . . . . . . . . . . . . . . . . . . 88
7
Chapter 1
Introduction
Malware, short for malicious software, refers to software programs designed

to damage or do any kind of unwanted actions on a computer system such as
disrupting computer operations, gather sensitive information, bypass access
controls, gain access to private computer systems and display unwanted ad-
vertising. Malware can be divided into the following categories not mutually
exclusive depending on their purpose.
• Adware. It is a type of malware that automatically delivers adver-

tisements. Advertising-supported software often comes bundled with
software and applications and most of them serve as a revenue tool.
• Spyware. It is a type of malware that spies and track user activity with-
out their knowledge. The capabilities of spyware can include keystrokes
collection, financial data harvesting or activity monitoring.
• Virus. A virus is a type of malicious software capable of copying itself

and spreading to other computers. Viruses can spread through email
attachments, through the network the computer is connected if any
other computer inside the network has been infected, by downloading
software from malicious sites, etc.
8
• Worm. It is a type of malware that they spread through the computer
network by exploiting operation system vulnerabilties. The major dif-
ference between worms and viruses is that computer worms have the
ability to self-replicate and spread independently while viruses rely on
human activity to spread.
• Trojan. A Trojan horse is a type of malware that disguises itself

as a normal file or program to trick users into downloading and in-
stalling malware. A trojan can give unauthorized access to the infected
computer and be used to steal data (logins, financial data, even elec-
tronic money), install more malware, modify files, monitor user activity
(screen watching, keylogging, etc), use the computer in botnets, and
anonymize internet activity by the attacker.
• Rootkit. A rootkit is a type of malicious software designed to remotely

access or control a computer without being detected by users or security
programs. For example, Rootkits can prevent a malicious process from
being visible or it can keep its files from being read, etc.
• Backdoors. A backdoor is a computer software designed to bypass

normal authentication procedures and compromise the system. Once a
system has been compromised, one or more backdoors may be installed
in order to allow access in the future without being detected by the user.
• Ransomware. It is a type of malicious software that essentially restricts

user access to the computer by encrypting the files or locking down the
system while demanding a ransom. Users are forced to pay the malware
author to remove the restrictions and gain access to their computer.
This type of payment is usually done with Bitcoins.
• Command & Control Bot. Bots are software programs created to au-
tomatically perform specific operations. Bots are commonly used for
DDoS attacks, spambots that render advertisements on websites, as
9
web spiders or for distributing malware. One way to defend against
bots is by using CAPTCHA tests in websites to verify users as human.
Nowadays, the detection of malicious software is done mainly with heuristic

and signature-based methods that struggle to keep up with malware evolu-
tion. Signature-based methods haven been heavily used for antivirus software
for decades. A malware signature is an algorithm or hash that uniquely iden-
tifies a specific virus. While identifying a particular virus is advantageous
it is quicker to detect a virus family through a generic signature. Virus re-
searchers found that all viruses in a family share common behaviors and a
single generic signature can be created for them. However, malware authors
always try to stay a step ahead of AV software by writing polymorphic and
metamorphic malware to do not match virus signatures. On one hand, poly-
morphic code uses a polymorphic engine to mutate while keeping the original
algorithm intact. Encryption is the most common way to hide code. On the
other hand, metamorphic viruses translate their own binary code into a tem-
porary representation, editing the temporary representation of themselves
and then translate the edited form back to machine code again. This muta-
tion can be done by using techniques such as changing what registers to use,
changing machine instructions to equivalent ones, inserting NOP instructions
or reordering independent instructions.
Therefore, AV vendors rely also in heuristic-based methods. This approach

is based on rules determined by experts which rely on dynamic analysis of
malicious behavior and thus, it can deal with unknown malware but it also
generates greater amounts of false positives than signature-based methods
because not each one of the detected suspicious executable files is a malware
file. As a result, AV vendors attempted to use an hybrid analysis approach
by using both signature-based and heuristic-based methods to tackle with
unknown malware. In contrast, before creating the signatures for malware,
it has first to be analyzed so to understand its capabilities and behavior. The
10
program capabilities and behavior can be observed either by examining its
code or by executing it in a safe environment.
1. Static Analysis. It refers to the analysis of a program without executing

it. The patterns detected in this kind of analysis include string signa-
ture, byte-sequence or opcodes frequency distribution, byte-sequence
n-grams or opcodes n-grams, API calls, structure of the disassembled
program, etc. The malicious program is usually unpacked and de-
cripted before doing static analysis by using disassembler or debugger
tools such as IDA Pro or OllyDbg which can be used to reverse com-
piled Windows executables and display malware code as a sequence of
Intel x86 assembly instructions.
2. Dynamic Analysis. It refers to the analysis of the behavior of a ma-

licious program while it is being executed in a controlled environment
(virtual machine, emulator, sandbox, etc). The behavior is monitored
by using tools like Process Monitor, Process Explorer, Wireshark or
Capture BAT. This kind of analysis tries to monitor function and API
calls, the network, the flow of information, etc. Compared to static
analysis, it is more effective and does not require the executable to be
disassembled but on the other hand, it takes more time and consumes
more resources than static analysis, being more difficult to scale. In
addition, as the controlled environment in which the malware is moni-
tored is different from the real one the program may behave different.
That’s because some behavior of malware might be triggered only un-
der certain conditions such as via a specific command or on a specific
system date and in consequence, can’t be detected in a virtual environ-
ment
In recent years, the possibility of success of machine learning approaches has

increased thanks to the confluence of three developments: (1) the rise of
commercial feeds that provide volumes of new malware, (2) the computing
11
1.1. OBJECTIVE
power has become cheaper meaning that researchers can fit large and more
complex models to data and (3) machine learning as discipline has evolved
and there are more tools at their disposal. Machine learning approaches hold
the promise that they might achieve high detection rates without the need
of human signature generation required by traditional approaches. In con-
sequence, AV companies and researchers begun to employ machine learning
classifiers to help them address this problem such as logistic regression[22],
neural networks[8] and decision trees[14].
The two principal tasks that have been carried out within the scope of mal-
ware analysis are (1) malware detection and (2) malware classification. First,
a file needs to be analyzed to detect if has any malicious content. In case
it exhibits any malicious content it is assigned to the most appropriate mal-
ware family according to their content and behavior through a classification
mechanism.
1.1 Objective
This master thesis aims to explore the problem of malware classification. In
particular, this thesis proposes two novel approaches based on Convolutional
Neural Networks (CNNs). On one hand, CNNs were applied for learning
discriminative patterns from malware images based on the work performed by
Nataraj et al.[21]. On the other hand, the CNN architecture proposed in [12]
was used to classify malicious software based on malware’s x86 instructions.
Both approaches have been evaluated on the data provided by Microsoft for
the BIG Cup 2015 (Big Data Innovators Gathering).
12
1.2. ORGANIZATION
1.2 Organization
The thesis is organized following chapters. The first and current chapter is
the introduction, which also contains the objectives and the organization of
the thesis. The second chapter introduces the background of the project,
focusing on neural networks and deep learning from its beginning until now.
The third chapter presents the state of the art review with special attention
on the machine learning algorithms and features used to detect and classify
malware. The fourth chapter introduces the Kaggle platform and the Mi-
crosoft’s Malware Classification Challenge. In addition, it also describes two
solutions of the competition. The fifth chapter describes the approach based
on the representation of malware as gray-scale images and the sixth chapter
explains how Convolutional Neural Networks can be used to extract features
from malware’s x86 instructions represented via word embeddings. Finally,
the last chapter wraps up the conclusions and the future work to be done.
13
Chapter 2
Background
Nowadays Deep Learning is the hottest topic in the Artificial Intelligence

and Machine Learning field. Deep Learning refers to the set of techniques
used for learning in Neural Networks with many layers. However, it is based
on a set of previous ideas that appeared over the 60s. In this chapter are
explained the most important concepts behind deep learning.
2.1 Artificial Neural Networks

Artificial Neural Networks are a family of models inspired by the way biolog-
ical nervous systems, such as the brain, process information which enables a
computer to learn from data. There are different types of NNs but in this
thesis are only presented two of them: (1) feed-forward networks and (2)
convolutional neural networks. First it is introduced the architecture of a
feed-forward network. Thus, this type of nets are composed by at least three
layers, one input layer, one output layer and one or more hidden layers. A
feed-forward network has the following characteristics:
• Neurons are arranged in layers, with the first layer taking in inputs and
the last layer producing the output.
14
2.1. ARTIFICIAL NEURAL NETWORKS
• Each neuron in one layer is connected to every neurons in the next

layer.
• There is no connection between neurons in the same layer.
To understand how a neural network works, first it has to be understood

what are neurons, how neurons learn and how the information pass from one
neuron to another.
2.1.1 Perceptrons
A perceptron was the earliest supervised learning algorithm and it is the
basic building block of Artificial Neural Networks (ANN). It was first intro-
duced in 1957 at the Cornell Aeronautical Laboratory by Frank Rosenblatt.
It works by taking several inputs (x1 , x2 , ..., xj ) and producing a single output
(y). Rosenblat introduced weights (w1 , w2 , ..., wj ) to express the importance
of the respective inputs to the output. The output of the perceptron is either
q
0 or 1 and it is determined by whether the weighted sum j wj ú xj + b is
less than or greater than 0.
Y Z
_
_
_
]
1: if w ú x + b > 0 _
_
_
^
output =
_
_ _
_
_
[0 _
: if w ú x + b <= 0\
However, a perceptron can only output zero or one, making impossible to

extend the model to work on classification tasks with multiple categories.
This issue can be solved by having multiple perceptrons in a layer, such that
all these perceptrons receive the same input and each one is responsible for
one output of the function. In fact, Artificial Neural Networks (ANNs) are
nothing more than layers of perceptrons (neurons or units as called nowa-
days). A perceptron can be seen as an ANN with only one layer, the output
layer, with 1 neuron.
15
2.1.2 Sigmoid neuron

The main limitation of perceptrons is that there are very difficult to tune,
because minimum changes in the weights and bias of any single perceptron
can cause the output to change drastically by completely flip, from 0 to 1 or
viceversa. And if we have a network of perceptrons, a single flip can com-
pletely change the behavior of the rest of the network.
This problem was solved by the introduction of the sigmoid neuron.
Exactly as the perceptron, a sigmoid neuron has inputs (x1 , x2 , ..., xj ) and
it also has weights for each input and a bias, but the output can be a real
number. The sigmoid function is defined as:
1
‡(z) =
1 + e≠z
And the output of the sigmoid neuron is:
1
q
1 + exp(≠ j wj ú xj ≠ b)
Hence, the only difference between the perceptron and the sigmoid neuron is
the activation function.
2.1.3 Loss function

To measure the performance of the neural network it is defined a function,
typically named cost or loss function which given a prediction or set of pre-
dictions and a label or a set of labels measures the discrepancy between the
algorithms prediction and the correct label. There are various cost functions
but the most common and simple in neural networks is the mean squared
error (MSE).
16
The mean squared error can be defined as:

m
ÿ
L(W, b) = 1/m( ||h(xi ) ≠ y i ||2 )
i=1
where:
• m is the number of training examples
• xi is the ith training sample
• y i is the class label for the ith training sample
• h(xi ) is the algorithm’s prediction for the ith training sample
The goal in training neural networks is to find weights and biases that min-
imizes some cost/loss function C. For that, it is used an algorithm called
gradient descent.
2.1.4 Gradient Descent Algorithm

Gradient Descent is an algorithm for minimizing the loss function. It is used
to find the local minimum of the loss function. Next you will find an outline
of the algorithm.
1. Start with a random initialization of each weight and bias in the NN. It
is important to randomly initialize all parameters because if not, if all
parameters start off at identical values, then all the hidden layer units
will end up learning the same function of the input. In consequence,
random initialization serves the purpose of symmetry breaking.
2. Keep iterating to update the parameters W,b as follows until it hope-

fully ends up at a minimum:
ˆ
l
Wi,j = Wi,j
l
≠– l
L(W, b)
ˆWi,j
17
ˆ
bli = bli ≠ – L(W, b)
ˆbli
where – is the learning rate and Wi,j l
and bli denote each weight and
bias in a particular layer l in the NN, respectively.
The derivative of the overall loss function can be computed as:
ˆ 1 ÿm
ˆ
l
L(W, b) = [ l
L(W, b : xi , y i )]
ˆWi,j m i=1 ˆWi,j
ˆ 1 ÿm
ˆ
l
L(W, b) = [ L(W, b : xi , y i )]
ˆbi m i=1 ˆbli
The learning rate is used to control how big a step is taken downhill with
gradient descent. Selecting the correct learning rate is critical. On one hand,
if – is too small, gradient descent can be slow. On the other hand, if – is
too large, gradient descent can overstep the minimum and even diverge.
Figure 2.1: Effects of different learning rates
18
2.1.5 Backpropagation
The key step is to compute all those partial derivates presented before. There-
fore, to compute efficiently these partial derivates is used the backpropagation
algorithm.
The intuition behind is as follows. Given a training example (xi , y i ) first

of all it is ran a forward pass to compute all the activations through the
network (also the output value of the hypothesis h(xi )). Thus, for each node
i in layer l it is computed an error term, ”il , that measures how much that
node was responsible for any errors in the output. For an output node, the
error term is measured directly from the difference between the network’s
activation and the true target value and is defined as ”inl , where nl is the out-
put layer. For hidden units, it is computed ”il based on the weighted average
of the error terms of the nodes that use ali as an input (ali = activation of
node i in layer l).
Overview of the algorithm:
1. Compute the activation of the layers l1 , l2 and so on up to the output

layer lnl by performing a feedforward pass.
2. For each output unit i in layer nl compute the error term ”inl .
ˆ 1 2
”inl = nl ú ||y ≠ hW,b (x)||
ˆzi 2
3. For all hidden layers l = nl ≠ 1, nl ≠ 2, ..., 2. Compute for each node i

in layer l: Q R
l+1
ÿ
”il = a Wjil ”jl+1 b ú f (zil )
Õ
j=1
19
4. Finally, compute the desired derivatives given as:
ˆ
l
L(W, b : x, y) = alj ú ”il+1
ˆWi,j
ˆ
L(W, b : x, y) = ”il+1
ˆbli
20
2.2. CONVOLUTIONAL NEURAL NETWORKS
2.2 Convolutional Neural Networks

A Convolutional Neural Network (CNN) is a type of feed-forward NN in
which the connectivity pattern between its neurons is inspired by the orga-
nization of the animal visual cortex, whose individual neurons are arranged
in such a way that they respond to overlapping regions tilling the visual field.
CNN are composed by three types of layers: (1) fully-connected, (2) con-
volutional (3) and pooling. All the various implementations of CNN can be
loosely described as involving the following process:
1. Convolve several small filters on the input image.
2. Subsample this space of filter activations.
3. Repeat steps 1 and 2 until you are left with enough high level features.
4. Apply a standard feed-forward NN to the resulting features.
Figure 2.2: AlexNet architecture
Figure 2.2 corresponds to the architecture used in [16] that was applied to
the ImageNet classification contest. The architecture consists of 8 learnable
layers, the first five are convolutional and the rest are fully-connected layers.
21
2.2.1 Local connectivity

It is impractical to connect neurons to all neurons in the previous layer when
dealing with high-dimensional inputs like images because in such network
architectures the spatial structure of the data is not taken into account.
In convolutional neural networks each neuron is connected to a small re-
gion of the input neurons (each neuron connects only to a small contiguous
region of pixels in the input) and thus, CNNs are able to exploit spatially
local correlation by enforcing a local connectivity pattern between neurons
of adjacent layers.
2.2.2 Convolutional Layer

Convolutional layers are the core of a CNN. A convolutional layer consists of a
set of learnable kernels which are convolved across the width and the height
of the input features during the forward pass producing a 2-dimensional
activation map of the kernel.
As a summary, a kernel consists of a layer of connection weights with the
input being the size of a small 2D patch and the output being a single unit.
Figure 2.3: Convolution
In figure 2.3 it is represented how convolution works. Considering an im-

age of 5x5 pixels as in the figure, where 0 means values completely black and
255 means completely white. In the center of the figure, it has been defined
a kernel of 3x3 pixels with all eight 0 except one wright set at 1. The output
is the result of computing the kernel at each possible position in the image.
22
Whether or not the kernel is convolved through all positions it is determined

by the stride. For example, for stride 1, it outputs the typical convolution
but for stride 2 half of the convolutions are avoided because there should be
2 pixels of distance between centers.
The size of the output after convolving a kernel of size Z over an image
N with stride S is defined as:
N ≠Z
output = +1
S
2.2.3 Pooling Layer

Pooling is a form of non-linear down-sampling. There are several non-linear
functions to implement pooling such as the minimum, the maximum and
the average but the most common is the maximum. Max pooling works by
partitioning the image into a set of non-overlapping rectangles and for each
sub-region outputs the maximum value.
Figure 2.4: Max pooling
The main benefits of max-pooling are:
1. It reduces computation for upper layers by eliminating non-maximal

values.
23
2. It provides a form of translation invariance and in consequence, pro-

vides additional robustness to position being a way of reducing the
dimensionality of intermediate representations.
24
2.3. OVERFITTING
2.3 Overfitting
Overfitting refers to the condition a predictive model describes the random
noise of a particular data instead of learning the underlying relationship. As
a result, these models may not yield accurate predictions for new observa-
tions.
This section describes the most common techniques used to avoid overfit-
ting in large networks.
2.3.1 Regularization
Regularization adds an extra term, named regularization term to the loss
function in a way that in consequence, the network would prefer to learn small
weights and penalize large weights. Regularization usually doesn’t affect
biases. That’s because large biases make it easier for neurons to saturate,
which is sometimes desirable. Moreover, having large biases doesn’t make a
neuron sensitive to its inputs in the same way as having large weights.
• L2 regularization.
⁄ ÿ 2
L(W, b) = L(W, b)0 + w
2n w
• L1 regularization.
⁄ÿ
L(W, b) = L(W, b)0 + |w|
n w
where L(W, b)0 is the original unregularized loss function.
25
2.3. OVERFITTING
2.3.2 Dropout
In large neural networks, it is difficult to average the predictions of different
networks at test time. To address this problem, dropout was introduced in
[32] by Geoffrey Hinton. The idea behind is to randomly drop units (along
with their connections) from the neural network during training to prevent
neurons from co-adapting too much.
At each iteration, we randomly and temporaly disconnect a percentage of

the hidden neurons (usually, half of them) in the network except those in
the input and the output layer. After that, the input is propagated through
the modified network and then the result is backpropagated also through the
modified network. After updating the appropiate weights and biases, the
process is repeated by, first restoring the dropout neurons, then choosing a
new random subset of hidden neurons to disconnect and so on.
Heuristically, the result of dropout is like training different neural networks

and in consequence, the dropout procedure is like averaging the effects of a
large number of networks. These networks will overfit in different ways but
at the end, hopefully, the network effect of dropout will be to reduce over-
fitting. Lastly, by reducing neurons co-adaptations, a neuron cannot rely
on the presence of a particular other neuron and is forced to learn more ro-
bust features that are useful in conjunction with the other different random
subsets of neurons.
2.3.3 Artificially expanding the training data

One of the best ways of reducing overfitting is to increase the size of the
training data. Unfortunately, training data sometimes is expensive or difficult
to acquire. To attach this problem, training data can be artificially expanded.
In particular, in recent competitions such as CIFAR-10, training images were
26
2.3. OVERFITTING
cropped, flipped and transposed in various ways to expand the training data.
27
2.4. DEEP LEARNING
2.4 Deep Learning

As defined at the beginning of the chapter, deep learning is the term used
to describe the techniques used to learn in networks with many layers (deep
neural networks). What convert a neural network into a deep neural network
is basically the number of layers. However, if deep neural networks are just a
NN with more layers, why it has been until recently that they have attracted
that much attention? Well, it is mainly because of two reasons: (1) the
computational power needed to train this networks and (2) the vanishing
gradient problem.
1. The basic idea of software able to simulate the neocortex’s neurons

in an artificial neural network is decades old but because of the im-
provements in mathematical formulas and the increasingly powerful
computers, computer scientist have been able to model networks with
many more layers than before.
2. The vanishing gradient problem was introduced in [10] by Sepp Hochre-

iter which found that in a neural network with activation functions such
as the sigmoid or the hyperbolic tangent where the gradient range is
(≠1, 1) or [0, 1), backpropagation is computed by the chain rule, multi-
plying n of this small numbers from the output layer through a n-layer
network, meaning that gradient decreased exponentially with n. This
results in the front layers training much more slowly than other layers.
2.4.1 ReLU units

The vanishing gradient problem was not solved until 2010 by the introduction
of the Rectified Linear units (ReLU) [9]. The activation function of ReLU
units is defined as:
Œ
ÿ
f (x) = ‡(x ≠ i + 0.5) ¥ log(1 + ex )
i=1
28
2.4. DEEP LEARNING
The softmax function log(1 + ex ) can be approximated by the max function

or hard-max function f (x) = max(0, x).
Figure 2.5 presents a graphically comparison between the ReL and the sig-
moid function.
Figure 2.5: ReL and sigmoid functions comparison
The main difference between both functions is that:

• The sigmoid function has range [0, 1] while the ReL function has range
[0, Œ).
• The gradient of the sigmoid function vanishes as it is increased or de-

creased x while the gradient of the ReL function doesn’t vanish as x is
increased.
2.4.2 Gradient Descent Optimization Algorithms

There are three variants of the gradient descent algorithm, which differ in
how much data is used to compute the gradient of the objective function.
29
2.4. DEEP LEARNING
1. Batch Gradient Descent. It computes the gradients for the loss function
L(W,b) for the entire training set. It guarantees the convergence to the
global minimum for convex error surfaces and to a local minimum for
non-convex surfaces.
2. Stochastic Gradient Descent. It performs a parameter update for each

training example xi and label y i . It is used for online learning as it
performs one update at a time. It enables to jump to new and potential
better local minimum but complicates the convergence to the exact
local minimum.
3. Mini-batch Gradient Descent[17]. It computes the gradients for the

loss function L(W,b) only for a small batch of n training samples. Its
faster than batch gradient descent and leads to a better convergence
than stochastic gradient descent.
Next it is presented an outline of some algorithms used in deep learning to

optimize the gradient descent algorithm.
1. Momentum [25]
The simplest gradient algorithm known as steepest descent 2.1.4, mod-
ifies the weight at time step t according to:
ˆ
l
Wi,j = Wi,j
l
≠– l
L(W, b)
ˆWi,j
ˆ
bli = bli ≠ – L(W, b)
ˆbli
However, it is known that learning such scheme can be very slow. To
improve the speed of convergence of the gradient descent algorithm it
is included the momentum term in the formula:
ˆ
l
Wi,j,t+1 = Wi,j,t
l
≠– l
L(W, b) + “Wi,j,t≠1
l
ˆWi,j
30
2.4. DEEP LEARNING
ˆ
bli,t+1 = bli,t ≠ – l
L(W, b) + “bli,t
ˆbi
where “ is the momentum term. In consequence, the modification of
the weight vector at the step t depends on both the current gradient
and the weight change of the step t ≠ 1.
2. Adagrad [5]
Adagrad is an algorithm for gradient-based optimization that adapts
the learning rate to the parameters, performing smaller updates for
frequent parameters and larger updates for infrequent parameters.
Adagrad uses a different learning rate for every parameter Wi,j,t
l
at each
time step t. In its update rule, it modifies the general learning rate – at
each time step t for every parameter Wi,j,tl
based on the past gradients
that have been computed for Wi,j,t .
l
– ˆ
l
Wi,j,t+1 = Wi,j,t
l
≠ ú L(W, b)
Glt,ij +‘ l
ˆWi,j
where Glt,ij œ Rdxd is the diagonal matrix where each diagonal element
ij is the sum of the squares of the gradients Wi,j,t+1
l
up to time step t 24
and ‘ is smoothing term that avoids division by zero (¥ 1e ≠ 8).
3. Adam [13]
Adam is the acronym for Adaptive Moment Estimation. It is another
method that computes adaptive learning rates for each parameter. It
stores an exponentially decaying average of the past squared gradients
that we will denota vt and similar to momentum, it keeps an exponen-
tially decaying average of past gradients mt :
mt = —1 mt≠1 + (1 ≠ —1 ) ú gt vt = —2 vt≠1 + (1 ≠ —2 ) ú gt2
and gt is the gradient of the objective function and —1 and —2 are the
decay rates. mt and vt are the estimates of the first moment or mean
31
2.4. DEEP LEARNING
and the second moment or uncentered variance of the gradients, respec-

tively. mt and vt are initialized as vectors of 0’s and in consequence,
during initial time steps and when the decay rates are small they tend
to be biased towards 0. To solve this problem, they computed the
bias-corrected first and second moment estimates:
mt vt
m̂t = v̂ =
1 ≠ —1
t t
1 ≠ —2t
As a result, the Adam update rule is defined as:
–
Wt+1 = Wt ≠ Ô ú m̂t
v̂t + ‘
32
Chapter 3
State of the Art
During the last decade, researchers and anti-virus vendors have begun em-
ploying machine learning algorithms like the Association Rule, Support Vec-
tor Machines, Random Forests, Naive Bayes and Neural Networks to address
the problem of malicious software detection and classification. An overview
of the methods can be found in [26], [7] and [6]. Following a few of these
approaches used in literature are discussed.
1. Byte-sequence N-grams[33, 31]

The representation of a malware file as a sequence of hex values can be
effectively described through n-gram analysis. A N-gram is a contigu-
ous sequence of n hexadecimal values from a given malware file. Each
element in the byte sequence can take one out of 257 different values,
i.e. the 256 byte range plus the special ‘??’ symbol. Byte-sequence
N-grams were first presented in [33], where they proposed a N-gram
based algorithm for malware classification implemented in IBM’s an-
tivirus scanner. The algorithm used 3-grams as features and a neural
network as a classification model. Notice that the size of the features
increase exponentially being 2562 for a bigram model and 2563 for a tri-
gram model being the techniques for dimensionality reduction a must
in this kind of models.
33
2. Opcodes N-grams [11, 27, 2, 3, 30]
Similar to byte-sequence N-grams, n-gram models have been generated
from opcodes extracted from assembly language code files. An opcode
(abbreviated from operation code) is the portion of a machine language
instruction that specifies the operation to be performed. In particular,
[3] investigated the most frequent opcodes and the rare opcodes present
in both goodware and malware. The following two charts show the 14
most frequent opcodes in goodware and in malware, respectively.
Figure 3.1: Most frequent 14 opcodes for goodware
Figure 3.2: Most frequent 14 opcodes for malware
3. Portable Executable [37, 28]

PE format is a file format for executables, DLLs and object code com-
34
monly used in the Windows operating systems. PE format is a data
format that encapsulates the necessary information for the Windows
OS loader to manage the executable code. It includes information such
as dynamic library references for linking and API import and export
tables.
Features from Portable Executables (PE) are extracted by perform-

ing static analysis using structural information of PE and are useful
to indicate whether or not the file has been manipulated or infected
to perform malicious activity. In [37] they extracted the following fea-
tures from PE: (1) File pointers which denote the position within the
file as it is stored on disk; (2) Import sections which describe functions
from which DLLs were used and the list of DLLs of the executable
that are imported; (3) Exports section which describes the functions
that are exported; (4) Structure of the PE header such as features like
the code and file size, the creation time, etc. After the feature extrac-
tion process they build an ensemble of support vector machines. The
training set consisted of 9838 executables, from which 2320 were be-
nign executables, 1936 backdoors, 1769 spyware, 1911 trojans and 1902
worms. They tested the performance of the classifier on two datasets:
(1)Malfease and (2)VXheavens.
In [28], Invencea Labs build a Deep Neural Network to detect malware

by using a set of features derived from the numerical fields extracted
from the file binary’s PE packaging. The Deep NN consisted of three
layers: the input layer of 1024 input features and two hidden layers of
1024 PReLU units each one. The output layer consisted of one sigmoid
unit denoting whether the file is goodware or malware. In the following
picture you can see an outline of the framework.
35
Figure 3.3: Outline of Invencea’s Malware Detection Framework
4. Entropy [29, 18]

Malware authors use obfuscation techniques to pass through signature-
based detection systems of antivirus programs. For this reason, it is
examined the statistical variation in malware samples to identify packed
and encrypted samples. In [18] it was presented a tool to analyze the
entropy of each PE section in order to determine which executable
sections might be encrypted or packed. They found that the average
entropy is 4.347, 5.099, 6.801, 7.175 for plain files, native executables,
packed executables and encrypted executables, respectively.
5. API calls [4, 36]

API and function calls have been widely used to detect and classify
malicious software. An experiment was conducted to determine the
top maliciously used APIs. They retrieved the imports of all of the PE
files and proceeded to count the number of times each sample uniquely
imported an API. They found that there was a total of 120126 uniquely
imported APIs.
In [4] they proposed a malware detection system which uses Window
36
API call sequences. They used a 3rd order Markov chain, i.e. 4-grams,
to model the API calls. The malicious executables mainly consisted of
backdoors, worms and Trojan horses collected from VXHeavens. Their
detection system achieved an accuracy of 90%.
6. Use of registers
In [19], they proposed a method based on similarities of binaries behav-
iors. They assumed that the behavior of each binary can be represented
by the values of memory contents in its run-time. In other words, values
stored in different registers while malicious software is running can be
a distinguishing factor to set it apart from those of benign programs.
Then, the register values for each API call are extracted before and
after API is invoked. After that, they traced the changes of registers
values and created a vector for each of the values of EAX, EBX, EDX,
EDI, ESI and EBP registers. Finally, by comparing old and unseen
malware vectors they achieved an accuracy of 98% in unseen samples.
7. Call Graphs
A call graph is a directed graph that represents the relationships be-
tween subroutines in a computer program. In particular, each node
represents a procedure/function and each edge (f,g) indicates that pro-
cedure f call procedure g. This kind of analysis have been used for
malware classification with good results. In [15], they presented a
framework which builds a function call graph from the information
extracted from disassembled malware programs. For every node (i.e.
function) in the graph, they extracted attributes including library APIs
calls and how many I/O read operations have been been made by the
function. Then, they computed the similarity between any two mal-
ware instances.
37
8. Malware as an Image
In [21] a completely different approach to characterize and analyze ma-
licious software was presented. They represented a malware executable
as a binary string of zeros and ones. Then, the vector was reshaped into
a matrix and the malware file could be viewed as a gray-scale image.
They were based on the observation that for many malware families,
the images belonging to the same family appear to be very similar in
layout and texture.
Figure 3.4: Visualizing Malware as an Image
In addition, to compute texture features from malware images they

used GIST[34, 23]. For classification, they used k-nearest neighbors
with Euclidean distance and they obtained a classification rate of 0.9929
performing as state of the art results in the literature but at a signifi-
cantly less computational cost.
38
Chapter 4
Microsoft Malware
Classification Challenge
The content of this chapter is structured as follows. First of all, it is presented

the Microsoft Malware Classification Challenge and the platform where the
competition was hosted. Secondly, it is described the winner’s solution of the
challenge and the set of techniques they used to win the competition followed
by an approach that achieved a logloss of 0.0064 and used features extracted
from gray-scale images of malware. Lastly, it is presented the deep learning
library used to implement the neural networks.
4.1 What’s Kaggle?

Kaggle is a platform where a large community of data scientist comprised
from thousands of MsCs and PhDs from fields such as computer science,
statistics and maths compete to solve valuable problems. These problems
come from competitions that companies host in Kaggle or from competitions
that are part of class homework or projects in academic institutions.
39
4.2. MICROSOFT MALWARE CLASSIFICATION CHALLENGE
4.2 Microsoft Malware Classification Challenge

In 2015, Microsoft hosted a competition in Kaggle with the goal of classifying
malware into their respective families based on the their content and charac-
teristics. For this challenge, Microsoft provided a dataset of 21741 samples,
with 10868 for training and the other 10873 for testing, being a dataset of
almost half a terabyte uncompressed. Microsoft provided a set of malware
samples representing 9 different malware families. Each malware sample had
an Id, a 20 character hash value uniquely identifying the sample and a class,
an integer representing one of the 9 malware family names to which the mal-
ware belong: (1) Ramnit, (2) Lollipop, (3) Kelihos_ver3, (4) Vundo, (5)
Simda, (6) Tracur, (7) Kelihos_ver1, (8) Obfuscator.ACY, (9) Gatak. The
distribution of classes present in the training data is not uniform and the
number of instances of some families significantly outnumbers the instances
of other families.
Figure 4.1: Malware Classification Challenge: dataset
For each observation we were provided with a file containing the hexadec-
40
imal representation of the file’s binary content and a file containing metadata
information extracted from the binary content, such as function calls, strings,
sequence of instructions and registers used, etc, that was generated using the
IDA disassembler tool.
4.2.1 Bytes file

The bytes file is the raw hexadecimal representation of the malware’s binary
content. A snapshot of one of these bytes files is shown below.
Figure 4.2: Snapshot of one bytes file
Each record of the hexadecimal files is composed by:
• Byte Count. Two hex digits indicating the number of hex digits pairs
in the data field.
• Address. Four hex digits representing the 16-bit beginning memory

address offset of the data.
• Record Type. Two hex digits, 00 to 05, defining the meaning of the
data field.
• Data. A sequence of n bytes of data, represented by 2n hex digits.
41
• Checksum. Two hex digits, a computed value that can be used to verify
the record has no errors.
4.2.2 ASM file

The asm file, generated by the disassembler tool, is a log containing vari-
ous metadata such as rudimentary function calls, memory allocation, and
variable manipulation. A snapshot of one of these files is shown below.
Figure 4.3: Snapshot of one assembly code file
An assembly program is often divided into three sections:

1. The data section. It is used to declare initialized data or constants and
do not change at runtime.
2. The bss section. It is used for declaring variables. Contains uninitial-

ized data.
3. The text section. It keeps the actual code of the program.

Apart from the previous sections, an assembly program can contain other
sections such as:
• The rsrc section. It contains all the resources of the program.
• The rdata section. It holds the debug directory which stores the type,
size and location of various types of debug information stored in the
file.
42
• The idata section. It contains information about functions and data

that the program imports from DLLs.
• The edata section. It contains the list of the funcions and data that
the PE file exports for other programs.
• The reloc section. It holds a table of base relocations. A base relocation

is an adjustment to an instruction or initialized variable value that’s
needed if the loader couldn’t load the file where the linker assumed it
would.
Additionally, some other sections can appear as a result of applying a poly-

morphic or metamorphic techniques to hide the actual code.
The assembly language consists of three types of statements:
1. Instructions or assembly language statements. Are used to tell the

processor what to do. Instructions are entered one instruction per line.
Each instruction has the following format:
[ label ] mnemonic [ operands ] [ ; comment ]
where the fields in brackets are optional. A basic instruction has two
parts: (1) the name of the instruction or the mnemonic to be executed
(also known as opcodes); (2) the operands or the parameters of the
command.
INC COUNT ; Increment t h e memory v a r i a b l e COUNT
MOV TOTAL, 48 ; T r a n s f e r t h e v a l u e 48 i n t h e
; memory v a r i a b l e TOTAL
Next you will find the top 10 most used x86 instructions in the training
dataset.
43
Figure 4.4: Top 10 opcodes in the training dataset
In addition, the average of instructions per each malware family is the

following.
Figure 4.5: Average of opcodes per malware family
44
One particularity of the training dataset is that there are some mal-
ware samples that due to code obfuscation techniques do not have any
instruction.
#samples
Ramnit 0
Lollipop 2
Kelihos_ver3 4
Vundo 22
Simda 0
Tracur 0
Kelihos_ver1 6
Obfuscator.ACY 9
Gatak 0
Table 4.1: Number of samples per class with 0 instructions
2. Assembler directives or pseudo-ops. Are commands part of the assem-

bler syntax but are not related to the x86 processor instruction set. All
assembly directives start with a period (.).
. data
. bss
The .data and .bss directives change the current section to .data or
.bss, respectively.
3. Macros. Are basically a text substitution mechanism. A macro is a se-

quence of instructions, assigned by a name and could be used anywhere
in the program. The syntax for a macro definition is:
%macro macro_name number_of_params
<macro body>
%endmacro
45
4.3. WINNER’S SOLUTION
4.3 Winner’s solution

The competition was won by a team of three people, Jiwei Liu and Xueer
Chen from the University of Pittsburgh and Xiaozhou Wang from Redman
Technologies Inc. Their solution relied mainly in the extraction of the three
following features from malware.
1. Opcode 2,3 and 4-grams.
• They counted the frequent 1-gram opcodes by selecting only the

opcodes that appear more than 200 times in at least one asm file
and they ended up with 165 features.
• All possible 2-gram counts were included (27225 features)
• The 3-gram and 4-gram counts which were greater than 100 in at
least one asm file were also included (21285 and 22384 features,
respectively)
2. Segment line count. They counted the number of lines per section in
the asm file and they also counted the number of different sections in all
malware samples which curiously was 448 a number much greater than
9, the number of sections in which an asm is usually divided. That’s
because of the application of metamorphic and polymorphic techniques.
3. Asm file pixel intensity features. Instead of representing the bytes file
as pixels they read the asm file as a binary file. They found that the
first 800 pixel intensities were very useful features.
Then, they used XGBoost(a machine learning library focused on gradient

boosted trees) and cross-validation to find the subset of features that best
classified the malware in classes and they finally obtained the lowest public
and private logloss of 0.003082695 and 0.002833228, respectively. In the next
figure you can find the confusion matrix of the training data.
46
4.4. NOVEL FEATURE EXTRACTION, SELECTION AND FUSION
FOR EFFECTIVE MALWARE FAMILY CLASSIFICATION
Ramnit Lollipop Kelihos_ver3 Vundo Simda Tracur Kelihos_ver1 Obfuscator.ACY Gatak
Ramnit 1541 0 0 0 0 0 0 0 0
Lollipop 1 2476 0 0 0 1 0 0 0
Kelihos_ver3 0 0 2942 0 0 0 0 0 0
Vundo 0 0 0 475 0 0 0 0 0
Simda 2 0 0 0 39 1 0 0 0
Tracur 1 0 0 0 0 750 0 0 0
Kelihos_ver1 0 0 0 0 0 0 398 0 0
Obfuscator.ACY 0 0 1 0 0 0 0 1225 2
Gatak 0 1 0 0 0 0 0 5 1007
Table 4.2: Winner’s solution: confusion matrix
where they classified correctly 10854 of 10868 samples (0, 9987%).
4.4 Novel Feature Extraction, Selection and

Fusion for Effective Malware Family Clas-
sification
In [1] they presented an approach that extracts and combines different char-
acteristics from malware and their fusion according to a perclass weighting
paradigm. Their method achieved an accuracy of 0.998% on the Microsoft
Malware Challenge dataset but what is more interestingly about their ap-
proach is that they also extracted feature patterns from images of malware.
Next you will find the different extracted set of features.
1. N-Gram (1G and 2G):

1-Gram and 2-Gram features from the hexadecimal representation of bi-
nary files described with a 256-dimensional vector and 2562 -dimensional
vector for the 1-Gram and 2-Gram models, respectively.
2. Metadata (MD1 and MD2):

They extracted the size of the file and the address of the first byte
sequence.
47
3. Entropy (ENT):
Entropy can be defined as a measure of the amount of the disorder and
it is used to detect the presence of obfuscation in malware files and for
this reason they computed the entropy of all the bytes in a malware
file.
4. Image patterns (IMG1 and IMG2):

They extracted the Haralick features and the Local Binary Patterns
features from each malware sample represented as a gray-scale image.
5. String length (STR):

They extracted possible ASCII strings and its length from each PE
using its hex dump.
6. Symbol frequencies (SYM):

The frequencies of the symbols -, +, *, ], [, ?, @ are extracted from
the disassembled files because are typical of code designed to evade
detection by resorting to indirect calls or dynamic library loading.
7. Operation code (OPC):

They counted the number of times a subset of 98 operation codes ap-
peared in each disassembled file. These subset was selected based either
on their commonness or on their frequent use in malicious applications.
8. Register (REG):
They computed the frequency of use of the registers in x86 architecture.
9. Application Programming Interface (API):

They measured the frequency of use of the top 794 frequent APIs used
in malware binaries based on the analysis performed in https://www.
bnxnet.com/top-maliciously-used-apis/.
10. Section (SEC):

They extracted different characteristics from sections in the disassem-
48
bly files such as the total number of lines in .bss, .txt, .data, etc sections
or the proportion of lines in each section compared to the whole file.
11. Data Define (DP):

They computed the frequency of db, dw and dd instruction because
there are malware samples that do not contain any API call and only
contain few operation codes, because of packing.
12. Miscellaneous (MISC):

This features are composed by the frequency of 95 keywords manually
chosen from the disassembled code.
Following you will find a table containing the list of feature categories and
their evaluation with XGBoost.
Feature Category #Features Accuracy Logloss

ENT 203 0.9987 0.0155
1G 256 0.9948 0.0307
STR 116 0.9877 0.0589
IMG1 52 0.9718 0.1098
IMG2 108 0.9736 0.1230
MD1 2 0.8547 0.4043
MISC 95 0.9984 0.0095
OPC 93 0.9973 0.0146
SEC 25 0.9948 0.0217
REG 26 0.9932 0.0352
DP 24 0.9905 0.0391
API 796 0.9905 0.0400
SYM 8 0.9815 0.0947
MD2 2 0.7655 0.6290
Table 4.3: List of feature categories and their evaluation in XGBoost
After the feature extraction process, they combined the features using a
version of the forward stepwise selection algorithm. The original version of
this algorithm starts with a model containing no features and then gradually
49
increases the feature set by adding one feature at each iteration. Instead
of considering one feature at a time, they added all the subset of features
belonging to a feature category at a time, until when adding more features
didn’t increase the value of logloss. By combining the feature categories as
described earlier, they achieved a test logloss of 0.0063 positioning its solution
among the top 10 in the competition.
50
4.5. DEEP LEARNING FRAMEWORKS
4.5 Deep Learning Frameworks

Deep Learning is a hot field in Artificial Intelligence and Machine Learning,
and thus, there are various deep learning libraries available open-source. The
most popular are:
1. Caffe.
It is a python deep learning framework developed by the Berkeley Vi-
sion and Learning Center. It allows you to define if train using the
CPU or the GPU easily. Caffe benefits from having a huge repository
with pre-trained neural network models suited for many problems. It
has a great implementation for convolutional networks but it has no
implementation for recurrent networks.
2. Theano.
It is a python deep learning library which make use of symbolic graph
for programming the networks. It also allows you to visualize the com-
putation graphs with d3viz.
3. TensorFlow.
It is written with a Python API over a C/C++ engine that makes it
run fast. It is more than a deep learning framework, and it has tools to
support reinforcement learning and other algorithms. In addition, Ten-
sorFlow can also be deployed in phones thanks that it can be compiled
in ARM architectures.
4. Deeplearning4j.
It is a deep learning framework developed in Java. It aims to be the
scikit-learn library in the deep learning space.
5. Torch.
It is a computational framework written in Lua that supports machine
learning algorithms. It has been used by large scale companies such as
51
4.5. DEEP LEARNING FRAMEWORKS
Google and Facebook. However, it is not as well-documented as other

deep learning frameworks.
TensorFlow has been chosen mainly because it has a Python API, there’s
a lot of documentation available and it has a large community that it con-
tinuously develops the library. In addition, it is very easy to setup and to
learn and recently, they released TensorBoard, a tool to visualize TensorFlow
graphs and to plot some metrics such as the accuracy or the loss at each train-
ing iteration. Moreover, it provides support for distributed computing since
version 0.8 (currently 0.11).
52
Chapter 5
Learning Feature Extractors

from Malware Images
The problem of malware detection and classification is a very complex task

and there’s no perfect approach to tackle it. For this reason AV vendors rely
in hybrid approaches that make use of traditional signature-based, heuristic-
based and machine learning methods as well as human analysis.
This chapter presents a novel approach for malware classification based on

the work performed by Nataraj et al. [21] which introduced the idea of
representing malicious software as gray-scale images. Then, they extracted
different features using GIST and they used the k-nearest neighbor algorithm
for classification. Our approach differ in the point that we use Convolutional
Neural Networks for learning discriminative patterns from the malware im-
ages.
The next sections explain how malware can be visualized as images followed
by the architectures of the different CNNs tested and its specifications as
well as the results obtained in the Kaggle’s competition.
53
5.1. VISUALIZING MALWARE AS GRAY-SCALE IMAGES
5.1 Visualizing malware as gray-scale images

This thesis is highly motivated by the work in [21] which is based on the
observation that images of different malware samples from the same family
appear to be similar while images of malware samples belonging to a differ-
ent family are distinct. Moreover, if old malware is re-used to create new
malware binaries the resulting ones would be very similar visually.
In their work, they computed image based features to characterize malware.

For that purpose, to compute texture features they used GIST [24]. The
resulting feature vectors were used to train a K-nearest neighbor classifier
with Euclidean distance. As introduced in [21], a given malware binary file
can be read as a vector of 8 bit unsigned integers and organized into a 2D
array. Then, this array can be visualized as a gray scale image in the range
[0,255].
Figure 5.1: Visualizing Malware as a Gray-Scale Image
The main benefit of visualizing malware as an image is that the different

sections of a binary can be easily differentiated. In addition, as malware au-
thors only change a small part of the code to produce new variants, images
are useful to detect small changes while retain the global structure. In con-
sequence, malware variants belonging to the same family appear to be very
similar as images while also being distinct from images of other families.
54
5.1.1 Malware families

Following are presented some malware files of each malware variant in the
dataset. One particularity of the dataset is that the samples do not contain
the PE header because it was removed to ensure sterility.
1. Ramnit. This type of malware is known to steal your sensitive infor-

mation such as user names and passwords and it also can give access
to an illegitimate user to your computer.
Figure 5.2: Rammit samples
2. Lollipop. This malware shows ads in your browser and redirects your
search engine results. In addition, it tracks what you are doing on your
computer. This type of malware usually is downloaded from the pro-
gram’s website or by some third-party software installation programs.
Figure 5.3: Lollipop samples
3. Kelihos_ver3. Third version of the Kelihos botnet. Kelihos is mainly

involved in spamming and theft of bitcoins. This trojan can give ac-
cess and control of your computer to an illegitimate user and can also
communicate with other computers about sending spam emails, run
malicious programs and steal sensitive information.
55
Figure 5.4: Kelihos_ver3 samples
4. Vundo. This trojan is known to cause popups and advertising for rogue
antispyware programs. In addition, sometimes is used to perform denial
of service attacks and also to deliver malware to other computers.
Figure 5.5: Vundo samples
5. Simda. It is a family of backdoors that try to steal sensitive information

such as usernames, passwords and certificates via its keylogging and
HTML injection routines. It also can give a hacker access to your
computer.
Figure 5.6: Simda samples
6. Tracur. This trojan hijacks results from different search engines such
as google, youtube, yahoo, etc, and redirects to a different web page. It
also give a hacker access to your computer and can be used to download
other types of malware.
56
Figure 5.7: Tracur samples
7. Kelihos_ver1. First version of the Kelihos botnet. It was first discov-

ered at the end of 2010 having infected 45.000 machines and sending
about 4 billions spam messages per day.
Figure 5.8: Kelihos_ver1 samples
8. Obfuscator.ACY. This class comprises all malware that has been ob-
fuscated to hide their purposes and to not be detected. The malware
that lies underneath this obfuscation can have almost any purpose.
Figure 5.9: Obfuscator.DCY samples
9. Gatak. It is a trojan that gathers information about your pc and sends

it to a hacker. It also downloads other malware files in your computer.
This trojan is usually downloaded when downloading a key generator
or a software crack.
57
Figure 5.10: Gatak samples
58
5.2. CNN ARCHITECTURES
5.2 CNN Architectures

In this thesis, we have proposed a novel approach to classify samples of
malicious software from their representation as gray-scale images. In the
work of [21] they used a traditional recognition approach to classify gray-
scale images of malware. First of all they extracted texture features from the
malware gray-scale images and then, they trained a K-NN classifier.
Figure 5.11: Shallow Approach
The main problem of their approach is that it doesn’t scales well with lots
of data. Accordingly, two ways of improvement are (1) keep building more
features like SIFT, HoG, etc and (2) using another classifier like Random
Forests or SVM. Instead, our approach makes use of Convolutional Neural
Networks to learn a feature hierarchy all the way from pixels to the layers of
the classifier.
This section presents the different architectures of the network and its spec-
ifications. The details of the architectures are defined in figures 5.12, 5.13
and 5.14.
All architectures have in common the input and the output layers. On one
hand, the input layer consists of N neurons, being N the size of the training
images. The image and the height of the images varies depending on the file
59
size and thus, before feeding the images as input all images had been down-
sampled to 32 by 32 pixels. In consequence, N is equals to 32ú32 = 1024. On
the other hand, all architectures have an output layer of 9 neurons because
the architectures are designed to handle a 9-class classification problem. In
addition, after each densely-connected layer it was applied dropout to reduce
overffiting.
To determine the parameters for each architecture it was performed a grid

search. Specifically, the grid search was used to determine the optimum learn-
ing rate, the size of the kernels of each convolutional layer and the number
of filters applied and also the number of neurons in each densely-connected
layer. Finally, to reduce the search space some parameters were fixed such as
the mini-batch size to 256, the region of the max-pooling layer to 2x2 with
stride equals to 2 and the learning rate to 0.001
60
5.2.1 CNN A: 1C 1D
The architecture consists of:
1. Input layer of NxN pixels (N=32).
2. Convolutional layer (64 filter maps of size 11x11).
3. Max-pooling layer.
4. Densely-connected layer (4096 neurons)
5. Output layer. 9 neurons.
The input layer consists of 32x32 neurons and is followed by a convolutional

layer composed by 64 filters of size 11x11. The output of the convolutional
layer is (32 ≠ 11 + 1) ú (32 ≠ 11 + 1) = 22 ú 22 = 484 for each feature map. As a
result, the total output of the convolutional layer is 22ú22ú64 = 30976. After
that, the pooling layer takes the output of each feature map from the con-
volutional layer and outputs the maximum activation of all 2x2 regions. In
consequence, the output of the pooling layer is reduced to 11 ú 11 ú 64 = 7744.
The pooling layer is then followed by a fully-connected layer with 4096 neu-
rons and every neuron of this layer is also connected to each one of the
neurons in the output layer.
The number of learnable parameters P of this network is:
P = 1024ú(11ú11ú64)+64+(11ú11ú64)ú4096+4096+4096ú9+9 = 39690313
where (11 ú 11 ú 64) + 64 are the shared weights for every feature map and
64 is the total number of shared bias.
61
Figure 5.12: Overview architecture A: 1C 1D
5.2.2 CNN B: 2C 1D
6. Densely-connected layer (512 neurons).
As in the previous architecture, the input layer consists of 32x32 neurons and
is followed by a convolutional layer composed by 64 filters of size 3x3. The
62
output of the convolutional layer is (32 ≠ 3 + 1) ú (32 ≠ 3 + 1) = 30 ú 30 = 900

for each feature map and a total of 30 ú 30 ú 64 = 57600. Next it is applied a
max-pooling layer which takes as input the output of the convolutional layer
and outputs the maximum activation of all 2x2 regions reducing the output
to 15 ú 15 ú 64. Then, the pooling layer is followed by another convolutional
layer of 128 filters with 3x3 receptive fields. After the convolutional layer
follows another pooling layer that takes as input the output of the previous
convolutional layer that is 13*13*128 and reduces the output to 7*7*128.
Finally, the polling layer is followed by a densely-connected layer with 512
neurons.
P = 1024 ú (3 ú 3 ú 64) + 64 + (15 ú 15 ú 64) ú (3 ú 3 ú 128) + 128+
+(7 ú 7 ú 128) ú 512 + 512 + 512 ú 9 + 9 = 20395209
where (3 ú 3 ú 64) + 64 and (3 ú 3 ú 128) + 128 are the shared weights for every
feature map and 64 and 128 are the number of shared bias in the first and
second convolutional layers,respectively.
63
Figure 5.13: Overview architecture B: 2C 1D
5.2.3 CNN C: 3C 2D
64
It starts with an input layer with 32x32 neurons which is then followed by a
convolutional layer with 64 filters of size 3x3. The output of the convolutional
layer is 30x30x64 and is used to feed the following max-pooling layer that
reduces its input to 15x15x64. Next follows the second convolutional layer
with 128 filters of size 3x3. After the convolutional layer it follows the second
pooling layer that takes as input the output of the second convolutional layer
(13 ú 13 ú 128) and outputs 128 feature maps of size 7x7. Moreover, a third
convolutional layer with 256 filters of size 3x3 follows the second pooling layer
which outputs 256 feature maps of size 5x5. Additionally, a third pooling
layer follows the convolutional layer reducing the input to 256 feature maps
of size 3x3. Lastly, follows two densely-connected layers of 1024 and 512
neurons, respectively.
P = 1024ú(3ú3ú64)+64+(15ú15ú64)ú(3ú3ú128)+128+(7ú7ú128)ú(3ú3ú256)+
+256 + (3 ú 3 ú 256) ú 1024 + 1024 + 1024 ú 512 + 512 + 512 ú 9 + 9 = 34519497
where (3 ú 3 ú 64) + 64, (3 ú 3 ú 128) + 128 and (3 ú 3 ú 256) + 256 are the shared
weights for every feature map and 64 and 128 are the number of shared bias
of the first, second and third convolutional layers, respectively.
65
Figure 5.14: Overview architecture C: 3C 2D
66
5.3. RESULTS
5.3 Results
The content of this section is structured as follows. First are presented the
results of the CNNs obtained during training and validation and then, are
presented the scores achieved in the competition.
5.3.1 Evaluation
The dataset provided by Kaggle for training was divided into two:
1. The training set of size (N ≠ N/10) = 9781
2. The validation set of size M = N/10 = 1086
where N is the total size of the dataset, N = 10868 and M = 1086. The
validation set was used to search the parameters of the networks and to know
when to stop training. In particular, we stopped training the network if the
validation loss increased in 10 iterations.
The next figure shows the accuracy and the cross-entropy achieved by the
models presented in 5.2 until they reached the 100th training iteration.
67
5.3. RESULTS
(a) Training & Validation accuracy (b) Training & Validation Cross-Entropy
Figure 5.15: Approach A: CNNs training results
It can be observed that the performance of the CNN with only one con-
volutional layer performs poorly than the other nets. Next you will find the
performance of the networks on the training set at the 500th iteration.
• CNN 1C 1D. Accuracy: 0.9857 Cross-entropy: 0.0968

Rammit Lollipop Kelihos_ver3 Vundo Simda Tracur Kelihos_ver1 Obfuscator.ACY Gatak
Rammit 1534 0 0 0 1 0 0 1 5
Lollipop 0 2375 0 0 0 4 0 0 98
Kelihos_ver3 0 1 2937 0 0 0 0 0 4
Vundo 0 0 0 472 0 0 2 0 1
Simda 0 0 0 1 41 0 0 0 0
Tracur 2 0 0 2 0 737 0 4 10
Kelihos_ver1 0 0 0 0 0 0 387 0 11
Obfuscator.ACY 0 0 0 0 0 1 0 1219 8
Gatak 0 0 0 0 0 2 0 0 1011
Table 5.1: CNN 1C 1D: confusion matrix
68
5.3. RESULTS

Rammit 1539 0 0 0 0 0 0 1 1
Lollipop 0 2471 0 0 0 1 0 0 5
Kelihos_ver3 0 0 2938 0 0 0 4 0 0
Vundo 0 0 0 474 0 0 0 0 1
Simda 0 0 0 1 41 0 0 0 0
Tracur 0 0 0 0 0 750 0 0 1
Kelihos_ver1 0 0 0 0 0 0 394 0 4
Obfuscator.ACY 0 1 0 0 0 2 2 1223 0
Gatak 0 0 0 0 0 0 0 0 1013

Rammit 1533 0 0 0 0 1 0 5 2
Lollipop 0 2443 0 0 0 1 0 0 33
Kelihos_ver3 0 0 2938 0 0 0 4 0 0
Vundo 0 0 0 474 0 0 0 0 1
Simda 0 2 0 0 39 0 0 1 0
Tracur 0 0 0 0 0 747 1 0 3
Kelihos_ver1 0 0 0 0 0 0 393 0 4
Obfuscator.ACY 0 3 0 0 0 1 3 1211 3
Gatak 0 0 0 0 0 0 0 0 1013
It can be observed that the convolutional neural networks with one and three
convolutional layers had problems mainly while labeling samples from Ram-
mit, Lollipop, Tracur, Kelihos_ver1 and Obfuscator.ACY and they ended up
misclassifying some samples as belonging to the Gatak malware’s family. In
particular, the major number of misclassifications had been produced from
samples of the Lollipop family, with 98 and 33 incorrect classifications from
the convolutional net with one and three convolutional layers, respectively.
Moreover, it can be seen that the training error of the convolutional network
with two layers is lower than the other two because it greatly reduced the
number of samples misclassified as Gatak and it achieved a training accuracy
of 0.9978 very near to the one obtained by the winner’s solution (0.9987) and
a loss of 0.0231 which is also lower the obtained in [1] using only the subset of
features named IMG1 (Haralick features) and IMG2 (Local Binary Pattern
features) as represented in 4.3 which is 0.9718 & 0.1098 and 0.9736 & 0.1230,
69
5.3. RESULTS
respectively.
5.3.2 Testing
Usually, Kaggle provides a test set without label in their competitions and
the Microsoft Malware Classification Challenge is not different. Therefore,
to evaluate our models using the test set we have to submit a file with
the predicted probabilities for each class to Kaggle. These submissions are
evaluated using the multi-class logarithmic loss. The logarithmic loss metric
is defined as:
1 ÿ N ÿM
logloss = ≠ yi,j log(pi,j )
N i=1 j=1
where N is the number of observations, M is the number of class labels,
log is the natural logarithm, yi,j is 1 if the observation i is in class j and 0
otherwise, and pi,j is the predicted probability that observation i is in class
j.
This type of evaluation metric provides extreme punishment for being con-
fident and wrong. That is, if the algorithm makes a single prediction that
an observation is definitely true (1) when it is actually false, it adds infinity
to the error score making every other observation pointless. Hence, in their
competitions, Kaggle bound the predictions away from extremes by using
the following formula:
1 ÿN
logloss = ≠ (yi log(pi ) + (1 ≠ yi )log(1 ≠ pi ))
N i=1
Moreover, the submitted probabilities are not required to sum to one because
they are rescaled prior to being scored.
Additionally, submissions in Kaggle are evaluated with two scores, the public
score and the private score where the first one is calculated on approximately
70
5.3. RESULTS
30% of the test data and the second one is calculated on the other 70%.
(a) Public score (b) Private score
The best results where obtained by the CNN with 3 convolutional layers
and 2 densely-connected layers which obtained a public and a private score
at iteration 500 of 0.117629734 and 0.134821767, respectively. That is an
improvement of 94.64% and 93.86% with respect to the equal probability
benchmark (logloss=2.197224577) which is obtained by submitting 1/9 for
every prediction. In contrast, other models achieved their respective lowest
score between iteration 50 and 100 which coincide with the point where the
algorithm converges into a local minima but unfortunately they were not able
to learn a better underlying relationship on the training data and ended up
performing much worse than the convolutional network with 3 convolutional
layers.
71
Chapter 6
Convolutional Neural Networks

for Classification of Malware
Disassembly Files
As described in sections 4.3 and 4.4, the approaches that performed better in
the competition where those that extracted features from the disassembled
files such as n-grams counts. A n-gram is a contiguous sequence of n items
from a sentence. In our case, those items are opcodes extracted from the
disassembled files. However, the main problem that has extracting those n-
grams is that the number of features extracted increases exponentially as N
increases. In particular, a 2-gram model will result in a two-dimensional ma-
trix of size 2562 = 65536, a 3-gram model will result in a three-dimensional
matrix of 2563 = 16777216 features, a 4-gram model in a four-dimensional
matrix of 2564 = 4294967296 and so on which turns out to be very compu-
tationally expensive.
In this thesis, it is used the CNN architecture introduced by Yoon Kim

in [12] as an alternative to n-grams because is much more computationally
efficient.
72
The network trained by Yoon Kim was a simple CNN with one layer on
top of word vectors obtained using Word2Vec, an unsupervised neural lan-
guage model. These word vectors were trained using the model proposed
by Mikolov in [20] on 100 billion words of Google News avaliable in https:
//code.google.com/p/word2vec/.
Figure 6.1: Yoon Kim model architecture
The structure of this chapter is organized as follows. First, it is introduced

what are word embeddings and in particular, the Skip-Gram model. Then,
it is described the architecture of the CNN and lastly, the results obtained
are compared against the best solutions from Kaggle’s competition.
73
6.1. REPRESENTING OPCODES AS WORD EMBEDDINGS
6.1 Representing Opcodes as Word Embed-

dings
Traditionally, Natural Language Systems treated words as symbols, creating
encodings that are randomly generated and do not provide useful informa-
tion for the system regarding the relationship among symbols. For example,
"The" may be represented "Id156". By using vector representations of words
these problems are solved. There are various techniques of language model-
ing that allow to map words or phrases from vocabulary, let’s say the English
language, to vectors of real numbers in a low-dimensional space. These set of
language models differ in the way they learn vector representation of words
but are all based on the distributional hypothesis which considers that words
that are used and occur in the same contexts tend to share semantic meaning.
The approaches that are based on this hypothesis can be divided broadly
into two categories:
• Count-based methods. They compute how often some word co-occurs

with its neighbors in a corpus and then they map these count-based
statistics to a vector for each word.
• Predictive methods. They try to predict a word given its neighbors in

terms of learned small, dense embedding vectors.
Word2Vec [20] is a very efficient predictive model for learning word embed-
dings. Word2Vec comes with two different approaches to learn the vector
representations of words: (1) the Continuous Bag of Words (CBOW) and
(2) the Skip-Gram model. The main difference between both models is that
CBOW predicts target words from source context words while the skip-gram
model predicts source context words from target words (the context of a word
are the words to the left of the target and the words to the right of the target).
74
To learn the word embeddings we used the Skip-Gram approach and thus, an
explanation of the CBOW model is not provided because it is out of scope.
6.1.1 Skip-Gram model

The Skip-Gram model tries to predict each context word from its target word.
Thus, the input to the model is wi and the output is wi≠2 ,wi≠1 ,wi+1 ,wi+2 if
a window size of 2 is used.
Figure 6.2: Skip-gram model architecture
As the network can’t be feed with words just as text strings is needed a
way to represent words. For that purpose, first it is build a vocabulary of
words from the malware training samples. In the case all operation codes
appear in the samples the vocabulary will consist of 665 words. Accordingly,
a word like "push" is going to be represented as a one-hot vector. This vector
will have 665 components (one for every word in the vocabulary) and in the
position corresponding to the word "push" it will place a 1 and 0s in all of
75
the other positions.
The output layer depends on the window size. Thus, for a window size
one (just predicting one word to the left and to the right of the targets word)
the network will output a two-dimensional vector, with one dimension of the
vector containing the probabilities of the words in the vocabulary to appear
at the left of the target word and the other dimension containing the prob-
abilities of the words in the vocabulary to appear at the right of the target
word. The dimension of the hidden layer or embedding layer corresponds to
V ú E, where V is the size of the vocabulary and E is the embedding size.
The training objective of the Skip-gram model is to find word representations

that are useful for predicting surrounding words in a corpora. Formally, given
a sentence of words w1 , w2 , ..., wT the objective is to maximize the average
log probability defined as:
1ÿ T ÿ
log p(wt+j |wt )
T t=1 ≠cÆjÆc,j”=0
where c is the size of the training context (Larger c results in more training
examples and can lead to a higher accuracy at the expense of the training
time) and p(wo |wI ) is formulated as:
exp(vwo T vwI )
p(wo |wI ) = qW
w=1 exp(vw vwI )
ÕT
where vw and vw Õ are the "input" and "output" vector representations of words
in the vocabulary and W is the number of words in the vocabulary.
The main drawback of this formulation is that as W is often large (105 )

the cost of computing “log p(wt+j |wt ) is impractical. An alternative of the
full softmax is the Noise-Contrastive Estimation (NCE). NCE postulates
76
that if a model is able to differentiate data from noise by means of logistic

regression is a good model. NCE can be shown to approximately maximize
the log probability of the softmax. The basic idea behind NCE is to train
a logistic regression classifier to discriminate between samples from the data
distribution and samples from the "noise" distribution, based on the ratio of
probabilities under the model and the noise distribution. Under those circum-
stances, as the Skip-Gram model is only concerned with learning high-quality
vector representations the NCE can be simplified as:
K
ÿ
log ◊(vwo T vwI ) + Ewi ≥ Pn (w)[log ◊(≠vwo T vwI )]
i=1
which is used to replace every log(wo |wI ) term in the Skip-Gram objective.
In consequence, the task is to distinguish the target word wo from draws
of the noise distribution Pn (w) using logistic regression, where there are K
negative samples for each data sample (K ƒ 5 ≠ 20). The noise distribution
Pn (w) is a design parameter. It was selected the unigram distribution U (w)
of the training data as the noise distribution because it is known to work
well for training language models. This distribution assumes that each word
in a sequence is independent and thus, each value would be independent of
the other values. In consequence, we would need to estimate the probability
of a sequence S in the malware’s language model P(S|M). The probability
r
generated for a specific sequence is calculated as follows: P (S) = Sw P (w)
To find the word embeddings it was used a window size equals to 5, meaning
that for each target word, the skip-gram approach tried to predict the five
words to the left and to the right. Following you will find the visualization
of the learned embeddings, using the t-SNE algorithm. [35]
77
Figure 6.3: t-SNE representation of the word embeddings
For instance, the opcodes whose vector representations are most similar
to the opcode "push" are:
1. pop
2. insertps
3. fucomp
which makes sense because tons of push instructions in malware files are
followed by the pop instruction or viceversa and are the two opcodes most
used. To compute the similarity between two vectors p and q it was used the
Euclidean distance.
ı̂ n
ıÿ
d(p, q) = d(q, p) = Ù (q
i ≠ pi )2
i=1
78
6.2. CONVOLUTIONAL NEURAL NETWORK ARCHITECTURE
6.2 Convolutional Neural Network Architec-

ture
The input of the network is a window of N (default=10000) opcodes from
the disassembly file represented as one-hot vectors of size V. As the number
of opcodes in every file is different we just extracted the first 10000 opcodes
of each file. The criteria used to select N is the following:
• A lower N will result in not capturing enough information of malware

samples belonging to the classes Ramnit and Lollipop which the average
number of instructions per sample is greater than 20000.
• A higher N will result in an increase in the time needed for training

the network.
• 10000 is greater than the average of opcodes of almost all malware

samples except from those belonging to the Ramnit and Lollipop family
(figure 4.5).
In addition, all malware files with less than N opcodes will be filled with
UNKNOWN tokens ("UNK").
Regarding the embedding layer, is composed by a V x E matrix and it is

the responsible of mapping the opcodes into low-dimensional vector repre-
sentations. It’s essentially a lookup table that can be learned during training
or it can be initialized using the vector representations learned using any
language model such as the Skip-Gram approach explained in section 6.1.1.
79
Figure 6.4: CNN Embedding layer output
The embedding layer is followed by a convolutional layer which have K

(default 64) filters of different sizes (i x E, where i is an integer value rep-
resenting the number of opcodes it covers. Each filter slides over the whole
embedded input, but varies in how many opcodes it covers.
80
Figure 6.5: CNN Convolutional layer output
Following the convolutional layer it is performed max-pooling over the

output of a specific filter size and leaves us with an output of size K, which
is essentially a feature vector. Then, the output of the max-pooling layer
is combined into one long feature vector. Lastly, the features obtained in
the previous layer (with dropout applied) are used to generate the desired
predictions by doing a matrix multiplication and selecting the class with the
highest score.
81
Figure 6.6: CNN Max-pooling & output layer
82
6.3. RESULTS
6.3 Results
Next, it is detailed how the parameters of the network were selected and the
heuristic search that it was performed followed by the results obtained in
both the training and the test set.
6.3.1 Evaluation
Heuristic Search
The parameters of the network were found by performing an heuristic search.

First, it was selected the learning rate. After that, it was selected the em-
bedding size E followed by the number of filters K and the different filter
sizes. By default, the network was initialized with the following parameters:
1. Learning rate = 0.001
2. Embedding size = 32
3. #filters = 64
4. Filter sizes = [3,4,5]
5. Batch size = 64
Next you will find the values that were considered for each particular param-
eter of the network:
83
6.3. RESULTS
1. Learning rate: 0.01, 0.005, 0.001, 0.0005
(a) Accuracy Comparison (b) Cross-entropy Comparison
Figure 6.7: Heuristic Search: Learning Rate
It was selected a ⁄ = 0.001 because a large learning rate can make the
gradient descent to overstep the minimum and also it has been tested
in scientific publications that the value chosen works really well.
2. Embedding size: 8, 16, 32, 64
Figure 6.8: Heuristic Search: Embedding size
It can be observed in the plot that both the CNN trained with E=32
84
6.3. RESULTS
and E=64 performed better than the others while performing similar
but as a higher E implies a higher training cost it was decided to select
an embedding size equals to 32.
3. Number of filters: 32, 64, 100
Figure 6.9: Heuristic Search: #Filters
In this case, the accuracy and the cross-entropy is also practically

the same through all the training iterations for #filters=64 and #fil-
ters=100 and as the number of filters to learn has a direct relationship
with the training cost of the network it was decided to learn 64 filters
for each different filter size.
4. Filter sizes: [3,4,5], [2,3,4,5,6,7], [2,3,4,5,6,7,8]
85
6.3. RESULTS
Figure 6.10: Heuristic Search: Filter Sizes
Lastly, it was chosen to learn filters with size 2 to size 8 because as it

shown in the plot the results obtained are slightly better than the CNN
trained with the other set of filter sizes.
Based on the previous results, the parameters of the network were:
• Learning rate (–) = 0.001
• Embedding size (E) = 32
• #filters = 64
• filter sizes = [2,3,4,5,6,7,8]
After selecting the parameters it was trained the neural network until the
validation loss increased in 10 iterations continuously using a mini-batch size
equals to 256. However, for comparison purposes, it was decided to limit
the number of training epochs to 25. Notice that the number of training
iterations per epoch is 39. The number of training iterations per epoch is
computed as:
#iterations = N/batch_size
86
6.3. RESULTS
where N is the total size of the training dataset (N=9781). Next you will
find their respective confusion matrices and also the accuracy and the cross-
entropy of the models until epoch 25.
(a) Training & Validation Accuracy (b) Training & Validation Cross-Entropy
Figure 6.11: Approach B: CNNs training results
• CNN without pretrained word embeddings. Accuracy: 0.9952 Cross-

Entropy: 0.0203
Ramnit 1537 0 0 0 0 0 0 4 0
Lollipop 4 2471 0 2 0 0 0 1 0
Kelihos_ver3 0 0 2938 4 0 0 0 0 0
Vundo 0 0 0 474 0 0 0 1 0
Simda 0 0 0 0 42 2 0 0 1
Tracur 3 0 0 1 0 744 0 3 0
Kelihos_ver1 0 0 0 6 0 0 392 0 0
Obfuscator.ACY 11 0 0 9 0 0 0 1208 0
Gatak 0 0 0 0 0 0 0 3 1010
Table 6.1: CNN without pretrained word embeddings: confusion matrix
• CNN with pretrained word embeddings. Accuracy: 0.9947 Cross-Entropy:

0.0243
87
6.3. RESULTS

Ramnit 1528 0 0 8 0 1 0 4 0
Lollipop 2 2473 0 2 0 1 0 0 0
Kelihos_ver3 0 0 2938 4 0 0 0 0 0
Vundo 0 0 0 474 0 0 0 1 0
Simda 0 0 0 0 42 0 0 0 0
Tracur 4 0 0 1 0 746 0 0 0
Kelihos_ver1 0 0 0 6 0 0 392 0 0
Obfuscator.ACY 11 0 0 9 0 1 0 1207 0
Gatak 0 0 0 1 0 2 0 0 1010
Table 6.2: CNN with pretrained word embeddings: confusion matrix
Looking at the confusion matrices it can be observed that this approach do

not have the problem of misclassifying samples as belonging to the Gatak
family as the approach based on the representation of malware as gray-scale
images has. However, it has some difficulty in classifying:
1. samples of the Ramnit family which are incorrectly classified as belong-
ing to the Obfuscator.ACY and viceversa
2. samples of the Tracur family which are misclassified as Ramnit samples
3. samples of the Lollipop family which are incorrectly classified as be-

longing to the Ramnit or the Vundo family.
Moreover, the CNN with pretrained word embeddings has also misclassified
some Ramnit samples as Vundo.
6.3.2 Testing
The CNN models trained during 25 epochs were used to generate the prob-
abilities of each sample from the test set to belong to a malware family.
Following you will find the public and the private score achieved by each
model.
Public Private
without pretrained word embeddings 0.048533931 0.031669778
with pretrained word embeddings 0.048851643 0.036707683
Table 6.3: Approach B: Test scores
88
6.3. RESULTS
The CNN without pretrained word embeddings achieved an improvement

of 98,56% with respect to the equal probability benchmark while the CNN
with pretrained word embeddings achieved an improvement of 98,33%. Both
models are very close in terms of performance and no one can be declared
as better than the other. May the fact of generating the vectors represen-
tentations of words using samples of malware instead of using goodware is
the cause of this situation. Probably if the input of the Skip-Gram model
used to learn the word embeddings had been samples of goodware the results
would have been quite different. That’s because instructions such as xor or
nop are commonly used by malware author’s for obfuscation purposes. On
one hand the exclusive OR (XOR) operation is commonly used to obfuscate
particular sensitive strings in the code such as URLs or registry keys. This
type of obfuscation works like this:
1. The attacker chooses a 1-byte value to act as a key (range 0 -255).
2. Then malware’s code iterates through every byte of the data that needs
to be encoded, XOR’ing each byte with the selected key.
When the attacker needs to deobfuscate the string, it repeats the step #2
XOR’ing each byte in the encoded string with the key value.
On the other hand, the No operation (NOP) is used in a technique called

dead-code insertion which simply adds some ineffective and redundant in-
structions to change the appearance of code while keeping its behavior intact.
From my point of view, I think that a better representation of the words

in the vocabulary would have been achieved if goodware had been used as
the training data because the previous mentioned techniques and lots more
not explained in this thesis might produce unnecessary noise in the embed-
dings.
89
Chapter 7
Conclusions
This master thesis studies the problem of classifying malware into their corre-
sponding families. In order to explore the problem, we used one of the most
recent and biggest datasets publicly available which was provided by Mi-
crosoft for the BigData Innovators Gathering Cup (BIG 2015). This dataset
provides two files for each malware sample.
1. The hexadecimal representation of the file’s binary content.
2. The disassembled file generated by the IDA (Interactive DisAssembler)

tool which contains various metadata information extracted from the
binary content. Metadata information contains instructions and regis-
ters used by the malware, as well as the functions and data imported
from DLLs.
This thesis presents two novel and scalable approaches using Convolutional
Neural Networks to recognize the family a malware sample belongs.
• The first approach is motivated by [21] which introduced the idea of

representing the malware’s binary content as gray-scale images. The
main benefit of visualizing malware as an image is that the different
sections of the binary can be easily differentiated. Their work is based
90
on the observation that images of different malware samples from the
same family appear to be similar while images of malware samples
belonging to a different family are distinct. This property is useful to
classify new malware binaries that have been created by re-using old
malware. That’s because images are useful to detect small changes
while retaining the global structure and the new samples would be
very similar visually to the old ones. In consequence, in this thesis,
we studied the application of CNNs to learn a feature hierarchy all the
way from pixels to the layers of the classifier.
• The second approach uses the architecture introduced by Yoon Kim in

[12] for sentence classification but applied to a distinct domain. Instead
of classifying sentences from the English language we used their archi-
tecture to classify malware samples using the x86 instructions extracted
from the disassembled files. We trained two models, the first one with-
out pretrained word embeddings and the second one with pretrained
word embeddings generated using the Skip-Gram model.
The first and the second approach obtained a score of 0.134821767 and
0.031669778, respectively. That is an improvement of 93.86% and 98,56%
with respect to the equal probability benchmark (logloss=2.1972245577) which
is obtained by submitting 1/9 for every prediction. Unfortunately, neither
approach outperformed the winner’s solution of the competition which ob-
tained a logloss equal to 0.002833228. That’s because their solution combined
different features such as opcode 2,3 and 4-grams as well as the number of
lines per section in the disassembled files, among others. However, the re-
sults obtained are quite promising because both approaches are able to clas-
sify malware samples much faster than all those solutions that rely on the
manually extraction of features and thus, are more scalable.
91
7.1. FUTURE WORK
7.1 Future Work

Even that both approaches have been successfully applied, there is still a
huge margin of improvement. From now on, consider that the CNN used to
classify malware based on their representation as gray-scale images is named
CNN A and the CNN used to classify malware given the x86 instructions
of the disassembled file is named CNN B. Hence, one way to improve the
results is by merging the output of the last fully-connected layer of CNN A
and the output of the max-pooling layer of CNN B. Notice that the resulting
CNN should be feed with two types of data: (1) the pixels extracted from the
representation of malware as gray-scale images and (2) the opcodes extracted
from the disassembled files. The idea is somehow similar to how pretrained
word embeddings were load and used in the CNN trained on opcodes. First,
both models are trained independently. Therefore, CNN A will learn low-
level and high-level features from the images while CNN B will learn a set
of features (opcode patterns) from the disassembled files. Then, the learned
filters are loaded into the CNN that will combine both approaches and the
network is trained again to learn the weight matrix of the output layer. The
main advantage of combining both CNNs is that the resulting CNN might be
able to solve the problems of each particular approach. In other words, the
model may not has problems while classifying Ramnit samples as belonging
to the Gatak family and neither discriminating between the Obfuscator.ACY
and the Ramnit families.
Additionally, there is still a lot of work to be done regarding the approach

presented in section 6. One possible modification would be to increase the
number of input opcodes allowed from 10.000 to the maximum number of
opcodes in a file found in the entire dataset. By making this small modi-
fication, all the information about every malware sample will be captured.
In particular, this modification can help to improve the classification rate of
92
7.1. FUTURE WORK
samples belonging to the Ramnit and the Lollipop families. That’s because
samples from both families contain more than 20.000 opcodes per file in av-
erage.
Another possible modification is to expand the vocabulary by including the

registers, the data directives and the most common imported DLL functions
and API calls found in malware. That would be effective for classifying all
those samples that, because of applying code obfuscation techniques, do not
have any instruction. In contrast, by expanding the vocabulary, the number
of parameters to learn will increase, as well as the time needed for training
the CNN.
Finally, the word embeddings could be improved by using as training data

samples of executables belonging to goodware. The main idea is that there
are some of the disassembled files that have been obfuscated by applying
different techniques such as using the dead-code insertion technique or the
XOR operation, among others. In consequence, the word embeddings may
not be accurate enough and might not represent correctly the words in the
vocabulary.
93
Bibliography
[1] Mansour Ahmadi, Dmitry Ulyanov, Stanislav Semenov, Mikhail Trofi-

mov, and Giorgio Giacinto. Novel feature extraction, selection and fu-
sion for effective malware family classification. In Proceedings of the
Sixth ACM Conference on Data and Application Security and Privacy,
CODASPY ’16, pages 183–194, New York, NY, USA, 2016. ACM.
[2] Clint Feher Shlomi Dolev Asaf Shabtai, Robert Moskovitch and Yuval
Elovici. Detecting unknown malicious code by applying classification
techniques on opcode patterns. In Security Informatics. 2012.
[3] Daniel Billar. Opcodes as predictor for malware. International Journal

of Electronic Security and Digital Forensics, 1:156–168, 2007.
[4] R Manoharan Chandrasekar Ravi. Malware detection using windows

api sequence and machine learning. International Journal of Computer
Applications, 43, 2012.
[5] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient
methods for online learning and stochastic optimization. J. Mach. Learn.
Res., 12:2121–2159, July 2011.
[6] Ekta Gandotra, Divya Bansal, and Sanjeev Sofat. Malware analysis and
classification: A survey. Journal of Information Security, pages 56–64,
2014.
94
BIBLIOGRAPHY
[7] Dragos Gavrilut, Mihai Cimpoes, Dan Anton, and Liviu Ciortuz. Mal-
ware detection using machine learning. Proceedings of the International
Multiconference on Computer Science and Information Technology, page
735–741, 2009.
[8] Li Deng George E. Dahl, Jack W. Stokes and Dong Yu. Large-scale
malware classification using random projections and neural network.
ICASSP, 2013.
[9] Geoffrey E. Hinton. Rectified linear units improve restricted boltzmann

machines vinod nair. In Proceedings of the 27th International Conference
on Machine Learning (ICML),, pages 807 – 814, 2010.
[10] S. Hochreiter. Untersuchungen zu dynamischen neuronalen netzen.

diploma thesis, institut fur informatik, lehrstuhl prof. brauer, technische
universitat munchen, 1991.
[11] Javier Nieves Yoseba K. Penya Borja Sanz Igor Santos, Felix Brezo
and Carlos Laorden. Opcode-sequence-based malware detection. In
Engineering Secure Software and Systems, volume 5965.
[12] Yoon Kim. Convolutional neural networks for sentence classification.

CoRR, abs/1408.5882, 2014.
[13] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic
optimization. CoRR, abs/1412.6980, 2014.
[14] J.Z. Kolter and M.A. Maloof. Learning to detect and classify malicious
executables in the wild. Journal of Machine Learning Research, page
2721–2744, 2006.
[15] Deguang Kong and Guanhua Yan. Discriminant malware distance

learning on structural information for automated malware classification.
Technical report, SIGMETRICS’13, 2013.
95
BIBLIOGRAPHY
[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet

classification with deep convolutional neural networks. In Advances in
Neural Information Processing Systems, 2012.
[17] Mu Li, Tong Zhang, Yuqiang Chen, and Alexander J. Smola. Efficient
mini-batch training for stochastic optimization. In Proceedings of the
20th ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, KDD ’14, pages 661–670, New York, NY, USA, 2014.
ACM.
[18] Robert Lyda and James Hamrock. Using entropy analysis to find en-
crypted and packed malware. IEEE Security and Analysis, 5:40–45,
2007.
[19] Zahra Salehi Mahboobe Ghiasi, Ashkan Sami. Dynamic malware detec-
tion using registers values set analysis.
[20] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff
Dean. Distributed representations of words and phrases and their com-
positionality. In C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani,
and K. Q. Weinberger, editors, Advances in Neural Information Process-
ing Systems 26, pages 3111–3119. Curran Associates, Inc., 2013.
[21] L. Nataraj, S. Karthikeyan, G. Jacob, and B. S. Manjunath. Malware

images: Visualization and automatic classification, 2011.
[22] Anil Thomas Nikos Karampatziakis, Jack Stokes and Mady Marinescu.
Using file relationships in malware classification. Detection of Intrusions
and Malware, and Vulnerability Assessment, 7591:1–20, 2013.
[23] Aude Oliva and Antonio Torralba. Modeling the shape of the scene: A
holistic representation of the spatial envelope. International Journal of
Computer Vision, 42:145–175, 2001.
96
BIBLIOGRAPHY
[24] Aude Oliva and Antonio Torralba. Modeling the shape of the scene: A
holistic representation of the spatial envelope. Int. J. Comput. Vision,
42(3):145–175, May 2001.
[25] Ning Qian. On the momentum term in gradient descent learning algo-
rithms. Neural Netw., 12(1):145–151, January 1999.
[26] Smita Ranvee and Swapnaja Hiray. Comparative analysis of feature ex-
traction methods of malware detection. International Journal of Com-
puter Applications, 120, 2015.
[27] Clint Feher Nir Nissim Robert Moskovitch, Dima Stopel and Yuval
Elovici. Unknown malcode detection via text categorization and the
imbalance problem. IEEE International Conference on Intelligence and
Security Informatics, pages 156–161, 2008.
[28] Joshua Saxe and Konstantin Berlin. Deep neural network based malware
detection using two dimensional binary program features, 2015.
[29] Sjsu Scholarworks and Donabelle Bays. Structural entropy and meta-
morphic malware, 2013.
[30] R. Sekar, M. Bendre, D. Dhurjati, and P. Bollineni. A fast automaton-

based method for detecting anomalous program behaviors. In Proceed-
ings of the 2001 IEEE Symposium on Security and Privacy, SP ’01,
pages 144–, Washington, DC, USA, 2001. IEEE Computer Society.
[31] Ohm Sornil and Chatchai Liangboonprakong. Malware classification

using n-grams sequential pattern feature. International Journal of In-
formation Processing and Management, 4, 2013.
[32] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and
Ruslan Salakhutdinov. Dropout: A simple way to prevent neural net-
works from overfitting. J. Mach. Learn. Res., 15(1):1929–1958, January
2014.
97
BIBLIOGRAPHY
[33] Kephart J.O. Tesauro, G.J. and Gregory B Sorkin. Neural networks for
computer virus recognition. IEEE International Conference on Intelli-
gence and Security Informatics, 11, 1996.
[34] Antonio Torralba, Kevin P. Murphy, William T. Freeman, and Mark

Rubin. Context-based vision system for place and object recognition.
pages 273–280, 2003.
[35] L.J.P. van der Maaten and G.E. Hinton. Visualizing high-dimensional
data using t-sne. 2008.
[36] Nitin Rai Veeramani R. Windows api based malware detection and
framework analysis. International Journal of Scientific & Engineering
Research, 3, 2012.
[37] Yanfang Ye, Lifei Chen, Dingding Wang, Tao Li, Qingshan Jiang, and
Min Zhao. Sbmds: an interpretable string based malware detection
system using svm ensemble with bagging, 2009.
98

Convolutional Layer Implementation To Classify Malware in Banking Financial Services Industry

Uploaded by

Copyright:

Convolutional Layer Implementation To Classify Malware in Banking Financial Services Industry

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Convolutional Layer Implementation To Classify Malware in Banking Financial Services Industry

Uploaded by

Copyright:

UNIVERSITY OF INDONESIA

CONVOLUTIONAL LAYER IMPLEMENTATION TO CLASSIFY MALWARE IN

FACULTY OF COMPUTER SCIENCE

According to AV vendors malicious software has been growing exponentially

3 State of the Art 33

4 Microsoft Malware Classification Challenge 39

5 Learning Feature Extractors from Malware Images 53

6 Convolutional Neural Networks for Classification of Malware

2.1 Effects of different learning rates . . . . . . . . . . . . . . . . . 18

3.1 Most frequent 14 opcodes for goodware . . . . . . . . . . . . . 34

4.1 Malware Classification Challenge: dataset . . . . . . . . . . . 40

5.1 Visualizing Malware as a Gray-Scale Image . . . . . . . . . . . 54

5.7 Tracur samples . . . . . . . . . . . . . . . . . . . . . . . . . . 57

6.1 Yoon Kim model architecture . . . . . . . . . . . . . . . . . . 73

4.1 Number of samples per class with 0 instructions . . . . . . . . 45

5.1 CNN 1C 1D: confusion matrix . . . . . . . . . . . . . . . . . . 68

6.1 CNN without pretrained word embeddings: confusion matrix . 87

Malware, short for malicious software, refers to software programs designed

• Adware. It is a type of malware that automatically delivers adver-

• Virus. A virus is a type of malicious software capable of copying itself

• Trojan. A Trojan horse is a type of malware that disguises itself

• Rootkit. A rootkit is a type of malicious software designed to remotely

• Backdoors. A backdoor is a computer software designed to bypass

• Ransomware. It is a type of malicious software that essentially restricts

Nowadays, the detection of malicious software is done mainly with heuristic

Therefore, AV vendors rely also in heuristic-based methods. This approach

1. Static Analysis. It refers to the analysis of a program without executing

2. Dynamic Analysis. It refers to the analysis of the behavior of a ma-

In recent years, the possibility of success of machine learning approaches has

Nowadays Deep Learning is the hottest topic in the Artificial Intelligence

2.1 Artificial Neural Networks

• Each neuron in one layer is connected to every neurons in the next

• There is no connection between neurons in the same layer.

To understand how a neural network works, first it has to be understood

However, a perceptron can only output zero or one, making impossible to

2.1.2 Sigmoid neuron

And the output of the sigmoid neuron is:

2.1.3 Loss function

The mean squared error can be defined as:

• m is the number of training examples

• xi is the ith training sample

• y i is the class label for the ith training sample

• h(xi ) is the algorithm’s prediction for the ith training sample

2.1.4 Gradient Descent Algorithm

2. Keep iterating to update the parameters W,b as follows until it hope-

The derivative of the overall loss function can be computed as:

Figure 2.1: Effects of different learning rates

The intuition behind is as follows. Given a training example (xi , y i ) first

Overview of the algorithm:

1. Compute the activation of the layers l1 , l2 and so on up to the output

3. For all hidden layers l = nl ≠ 1, nl ≠ 2, ..., 2. Compute for each node i

4. Finally, compute the desired derivatives given as:

2.2 Convolutional Neural Networks

1. Convolve several small filters on the input image.

2. Subsample this space of filter activations.

4. Apply a standard feed-forward NN to the resulting features.

Figure 2.2: AlexNet architecture

2.2.1 Local connectivity

2.2.2 Convolutional Layer

Figure 2.3: Convolution

In figure 2.3 it is represented how convolution works. Considering an im-

Whether or not the kernel is convolved through all positions it is determined