AI-enabled Blockchain: An Outlier-aware Consensus Protocol

for Blockchain-based IoT Networks

Mehrdad Salimitari, Mohsen Joneidi, and Mainak Chatterjee
Department of Electrical Engineering and Computer Science
University of Central Florida, Orlando, FL 32816
Email: {mehrdad@cs, joneidi@ece, mainak@cs}.ucf.edu

Abstract—A new framework for a secure and robust consensus Using blockchain for the IoT devices in a smart home
in blockchain-based IoT networks is proposed using machine is non-trivial. This is primarily because the IoT devices are
learning. Hyperledger fabric, which is a blockchain platform resource-constrained and might not be able to perform the
arXiv:1906.08177v2 [cs.DC] 9 Aug 2019

developed as part of the Hyperledger project, though looks

very apt for IoT applications, has comparatively low tolerance extensive computations required to achieve consensus. For
for malicious activities in an untrustworthy environment. To example, blockchain consensus protocols such as Proof of
that end, we propose AI-enabled blockchain (AIBC) with a 2- Work is contingent upon solving compute-intensive hash func-
step consensus protocol that uses an outlier detection algorithm tions, having storage requirements, and needing low-latency
for consensus in an IoT network implemented on hyperledger communication links [7]. Resource-constrained devices cannot
fabric platform. The outlier-aware consensus protocol exploits a
supervised machine learning algorithm which detects anomaly ac- fulfill these requirements [8]. Recently, several novel con-
tivities via a learned detector in the first step. Then, the data goes sensus approaches have been proposed to overcome these
through the inherent Practical Byzantine Fault Tolerance (PBFT) limitations. One of the projects which tries to address these
consensus protocol in the hyperledger fabric for ledger update. challenges is the hyperledger project [6]– which is what we
We measure and report the performance of our framework with chose as our evaluation platform in this research.
respect to the various delay components. Results reveal that our
implemented AIBC network (2-step consensus protocol) improves We implement a blockchain-based IoT smart home net-
hyperledger fabric performance in terms of fault tolerance by work on hyperledger fabric which has low computational
marginally compromising the delay performance. requirements and fast network response time that makes it
Index Terms—Blockchain; Artificial intelligence (AI); Con- desirable for IoT applications [6], [8]. Hyperledger fabric uses
sensus protocol; Hyperledger fabric; Internet of Things (IoT); practical byzantine fault tolerance (PBFT) method which can
Outlier detection.
successfully reach consensus over new data if the ratio of
I. I NTRODUCTION malicious devices is less than 1/3 [9]. On the other hand, most
of the devices in a home network are resource constrained and
The wide spread deployment of IoT devices is enabling the
very vulnerable to different types of cyber-attacks. Therefore,
automation of various aspects in our daily lives. One of the
we should define a mechanism to detect malicious activities
success stories of IoTs has been the automation of homes and
and the compromised devices and disconnect them from the
cities which are referred as smart homes and smart cities. A
rest of the network in order to exclude them from participating
significant obstacle towards realizing the true vision of smart
in the PBFT consensus protocol.
homes is securing the communicated and stored data in the
In this paper, we propose the AI-enabled blockchain (AIBC)
home network against malicious acts desiring to wreak havoc
network with a 2-step consensus protocol using an outlier
with someone’s home [1].
detection algorithm as the first step and PBFT as the sec-
Though there are many competing technologies that try to
ond one. Outlier detection algorithm discovers anomalies in
immune data in smart homes against attacks, blockchain has
multimodal data that is captured by the various IoT devices
emerged as probably the most promising for both i) securing
in a smart home network. We use multimodal data fusion to
the home network against manipulation attacks on stored data
map the different data types received from different devices to
and ii) providing a secure platform for all the devices in
an intermediate domain. Since the captured data from different
the network to communicate with each other [2], [3]. In a
sensors are not independent of each other, the outlier detection
blockchain, the data is immutable because of the underlying
algorithm aims to learn the intrinsic structure of the data
consensus protocols– a process by which all transactions are
and exploits the inter-dependencies among data from different
validated by all the nodes [4]. Therefore, manipulation attacks
devices. Eventually, outlier data is rejected at the end of the
on transmitted or stored data are not plausible through a
first step and the corresponding device is refrained from going
single compromised node and majority of nodes ought to be
to the second consensus step (PBFT). Thus, the overall fault
compromised for a successful attack [5]. Different consensus
tolerance of the blockchain network is enhanced in comparison
protocols and their applicability towards IoT networks can be
to a naive hyperledger fabric implementation. Fig. 1 shows the
found in [1], [6].
intuition of our proposed 2-step consensus protocol.
This research was partially funded by Cyber Florida’s Collaborative Seed To the best of our knowledge, this is the first research
Award program. work that tries to practically apply machine learning to reach
 Uncompromised Data
A. Three-layer Architecture
Compromised Data The topology of our implemented 3-layer network is il-
Miss-detected Data lustrated in Fig. 2. The first layer is the application layer
X Approved Data containing n smart devices, smart meters, and sensors (A1
X PBFT Consensus X Rejected Data
to An ) within a smart home network. The second layer is the
Protocol False Alarm Data
X edge blockchain layer that includes m endorsing peers (Pe1
X to Pem ) and data aggregators. This layer endorses the new
Outlier-Aware
Blockchain
Network transactions (T x1 to T xn ) from applications and is partially
responsible for the 2-step consensus protocol. The third layer
IoT Network is the core blockchain layer consisting of an orderer and z
regular peers (P1 to Pz ). This layer creates a block out of
Fig. 1: The 2-step consensus protocol in our proposed AIBC
network. The attack ratio in this schematic is 37.5% which PBFT the received endorsed transactions (R1 /E1 to Rn /En , where
cannot tolerate in a conventional blockchain network. However, the R and E denote transaction results and their corresponding
AIBC network with outlier detector will attenuate the attack impact endorsements respectively) from the application layer and is
to 20% in the first step which can successfully go through PBFT as also responsible for the 2-step consensus protocol.
the second step consensus protocol.
There are n organizations (Org1 to Orgn ) in the AIBC
network each of which contains one application, at least
consensus in a blockchain network. Gupta et al. have suggested
one endorsing peer (n ≤ m), none or few regular peers
the possibility of applying machine learning techniques to
(n ≤ z or n ≥ z ). For ease of illustration, only one regular
blockchain’s consensus process as a future research work
peer and one endorsing peer are shown in each organization
without any investigation [10]. Dinh et al. have discussed the
in Fig. 2. Endorsing peers have an aggregator to receive
benefits of integration of blockchain and artificial intelligence
data (transactions) from the application layer, a copy of the
(AI) [11]. They suggest that a blockchain governed by a
chaincode (also known as smart contract, shown as C) to
machine learning algorithm might be able to detect attacks
endorse new transactions, a copy of the ledger (shown as L),
and invoke proper defence mechanisms or isolate the com-
and the detector (shown as Det) to participate in the 2-step
promised component. We have successfully formulated and
consensus protocol. Regular peers have only one copy of the
implemented this idea called AI-enabled blockchain (AIBC).
ledger and the detector for participation in the 2-step consensus
Dey has proposed a utility function similar to the function used
protocol to enhance the security of the blockchain network.
in [5] to detect anomaly [12]. Then, he claims this value can be
Peers (both endorsing and regular) are involved in the 2-step
used as a feed for a supervised machine learning algorithm to
consensus protocol. Thereby, as the number of peers (m and
measure the likeliness of an attack and prevent the blockchain
z) increases, more of them ought to be infected in order to
confirmation of that transaction in the consensus protocol.
prevent a successful consensus.
However, he does not propose an algorithm or implementation
to design a practical consensus protocol. B. Communication protocol in AIBC network
In order to test the validity of our 2-step consensus protocol
There exist two different communication protocols in the
for IoT networks, we implement a 3-layer architecture using
AIBC network: query process and invoke process.
hyperledger fabric. The first layer is the application layer
1) Query Process: In query process, an application con-
containing different IoT devices. The second and third layers
nects to an arbitrary endorsing or regular peer to get updated
are the edge blockchain layer and the core blockchain layer
about the current state of the ledger. This process requires
that contain different components of the AIBC.
only three steps: (i) the application connects to a peer directly
We measure the latency of different parts of our implemen-
(without an aggregator), (ii) the application sends query re-
tation and the number of outlier devices in each time unit
quest to the peer, and (iii) the peer accesses its copy of ledger
using synthetic data represented as a matrix. We compare
and sends the result back to the application.
our proposed architecture with the conventional hyperledger
fabric implementation and investigate the effect of the outlier
detection algorithm on fault tolerance. Results reveal that our
proposed AIBC network can reach consensus over new data Orderer
B Det P1 L Det P2 L Det Pz L
in milliseconds with better fault tolerance than a naive hyper-
ledger fabric implementation. This is achieved by detecting the
malicious devices in the first consensus step and prohibiting Det L Det L Det L C
C C
them from participating in the PBFT step. Pe1 Pe2 Pem
Aggregator Aggregator Aggregator
II. P ROPOSED S YSTEM A RCHITECTURE R1 E 1 Tx1 R2 E2 Tx2 Rn En Txn

In this section, we discuss the proposed 3-layer architecture A1 A2 An

R1 E1 R n En Org1 Org2 Orgn
and the communication protocol between the different com-
ponents of the implemented AIBC network.
Fig. 2: The topology of our implemented 3-layer AIBC network.
 5. Transactions sent to in the new block.
4. Order the new transactions
Orderer all peers in a block
Regular Peers 2-step Consensus Protocol: The two steps of the consensus
5.1(b) Update the relay switch protocol consist of the detector (which detects the outliers)
2.2 Chaincode temp. 5.2. Second step
consensus and execution of PBFT.
6. Peer updates ledger updates ledger
5. The orderer combines all the received transaction results and
5.1(a) Peer invokes detector
with new block
Det L Other
their corresponding endorsements in a new block and sends
1. Connect to endorsing peer C endorsing it to all the peers (both endorsing and regular) to initiate the
A1 3. Endorsed transaction result (R1/ E1) Pe1 2.1 Peer invokes
Peers 2-step consensus protocol.
2. Send new readings (Tx1) Aggregator chaincode Step 1: Detector.
7. Ledger update event Endorsing Peer 1 5.1(a) Each peer checks all the transactions within the block
Endorsing Peers
AIBC via its detector that uses the outlier detection algorithm
discussed in Section III to find outlier data and reject it.
Fig. 3: The invoke process communication protocol in the imple- 5.1(b) The detector excludes the peers associated with the
mented AIBC network using a relay switch. organization containing the application that generated the
outlier data from participating in the second step consensus. To
2) Invoke Process: During this process, an application do so, the detector updates the relay switch in order to notify
connects to its corresponding endorsing peers to request a its corresponding peer which set of peers (both endorsing and
change in the ledger by its new data. A simplified version regular) it must connect to for the second step consensus. This
of our implementation with just one application (A1 ) and its step can prevent more than 33.3% of compromised nodes to
corresponding endorsing peer (Pe1 ) is delineated in Fig. 3 to intervene in the PBFT consensus protocol to some extent–
clarify the communication protocol in an invoke process. It the exact fraction depends on the accuracy of the detector
should be noted that there are other applications, other endors- (discussed in Section III-C).
ing peers and regular peers as shown in Fig. 2. Other endorsing Step 2: PBFT.
peers communicate with their corresponding application and 5.2 Each peer verifies whether each of the transactions in the
simultaneously traverse all the communication steps shown in new block is endorsed by all the required peers specified in the
Fig. 3 (steps 1 to 7) whereas the regular peers go through endorsement policy. In addition, they check if the result of a
steps 5 and 6 only. The invoke process has three phases: (i) specific transaction is the same from all the required endorsing
the endorsement phase (denoted with solid lines), (ii) the 2- peers. This is to ensure that the application is not compromised
step consensus protocol (denoted with dashed lines) and (iii) and has not sent an incorrect result for the transaction. All the
the ledger update (denoted with dotted lines) in Fig. 3. transactions in the block are labeled as valid or invalid after
Endorsement Phase: This phase has 4 steps: verification of the endorsements by each of the peers. Then,
1. In this step, each application needs to connect to one or the peers connects to their trusted set of peers according to the
more endorsing peers according to the endorsement policy (not obtained relay switch in the previous step to reach a consensus
arbitrarily). Endorsement policy delineates which endorsing over the new block using PBFT method.
peers from which organizations are required to endorse a Ledger Update: After the completion of the consensus pro-
new transaction (T x) proposed by an application (A). In our tocol, the invoke process is finalized as:
implementation, each application communicates with only one 6. Each peer updates its copy of ledger.
endorsing peer as shown in Fig. 2. 7. Applications get notified about the ledger update.
2. The application (A1 ) reads from a smart device, and sends
its new readings to its corresponding endorsing peer (shown III. O UTLIER D ETECTION
as Pe1 ) in the AIBC network. In order to detect malicious devices, we propose to employ
2.1. The endorsing peer executes the new transaction (data) machine learning for outlier detection as the first step of
in its copy of chaincode to endorse it. consensus. To that aim, the low-rank assumption is considered
2.2. If the transaction get endorsed successfully, the chain- as the core model which is a popular assumption in machine
code temporarily updates that endorsing peer’s copy of ledger learning [13]. We implement the outlier detector using a
as a proposed update. secondary chaincode which is installed on all the peers in the
3. The endorsing peer sends the result of the proposed trans- blockchain network (shown by Det in Fig. 2). To apply the
action (R1 ) along with its endorsement (E1 ) obtained by outlier detection algorithm to the multi-modal data gathered
the endorsing peer’s digital signature to the corresponding from different sensors and devices, we need to first fuse the
application. data for a unified representation.
4. The application checks all the received results along with
their corresponding endorsements from corresponding endors- A. Multimodal Data Fusion
ing peers to have the same result with valid endorsements. Sensors and devices in a smart home deal with different
If these requirements are met, the application sends the result types of data. To impose a model for the ensemble of data in
with its corresponding endorsements to the orderer to be added AI-enabled systems, we need to map the data to a meaningful
intermediate domain [14]. The data from device n at time A straightforward criterion for rank estimation is based on the
slot t is denoted by dnt ∈ Rbn where bn is the dimension of Frobenius norm of residual.
received data from sensor n. Let dtP ∈ Rb represent the fused kD −
PR
λr ur v Tr kF
r=1
th
data at the t time slot where b = bn . Matrix D ∈ Rb×T Rest = min R s.t. ≤ . (3)
kDkF
represents collected data from N devices over T time slots.
In other words, D = [d1 · · · , dT ]. Here,  is a constant between 0 and 1. As  increases,
the required rank decreases. Let us define matrix Z as the
B. Outlier Detection Algorithm difference of D and D̃, i.e., Z = D − D̃. Each row of
For detecting inconsistent data, we use low-rank data struc- Z corresponds to the estimated perturbation of a device.
ture which is a typical assumption for many real-life data [15]. The threshold for margin of each device is defined by hn ,
This model is one of the most well-known data structures in which is a function of the desired false alarm probability of
signal processing and data mining [16]. However, considering detector. The parameters of the employed low rank model are
the recent advances in AI, deep learning methods and non- summarized in Table I. The learned detector is characterized
linear models can also be imposed [17], [18]. Without loss of by [U , Rest , h]. Vector h is concatenation of [h1 , · · · , hd ].
generality, the low-rank model is assumed in the present paper. The input data at time slot t, dt , must be analyzed according
Matrix D contains the training data to design the detector via to the learned detector by training data D. First, dt should be
rank decomposition model. The rank decomposition model can projected on the span of the learned detector as:
be constructed via singular value decomposition (SVD) of D d̃t = U (U T U )−1 U T dt , (4)
as follows,
rank
X where, d̃t is the projection of dt on the low-rank model.
D = U ΛV T = λr ur v Tr , (1) The residual of the projection is defined as the difference of
r=1
the measured data and the projected data on the model, i.e.,
where, diag(Λ) = [λi ] contains singular values of D. Trans- z t = dt − d̃t . This residual vector contains the mismatch of
pose of V is indicated by V T . Moreover, columns of U N devices for time t. The value of zt (n) is the metric for
and V are right singular vectors and left singular vectors, decision on detecting outliers, i.e., if |zt (n)| is greater than
respectively. The rank of D is upper bounded by Rmax = the threshold, hn , the measured data violates the margin of
min(T, b). Rank of a matrix is equal to the minimum number the model.
of rank-one components that holds Eq. (1). However, the actual
rank of a typical data could be much smaller than Rmax as the Algorithm 1 Outlier rejection based on low-rank model
measurements from devices are dependent on each other. The Require: training data (D), online time series dt and Pf a .
goal of the outlier detection algorithm is to learn the possible 1: compute U , Σ and V using Eq. (1).
dependencies and detect those patterns that do not agree with Initialization:
2: model.rank ← estimate using Eq. (3).
the rest of the measurements. By rejecting the inconsistent 3: model.bases ← U (:, 1 : model.rank).
data, the blockchain network becomes AI-enabled. In AIBC, 4: model.thresholds ← select thresholds based on Pf a .
the devices associated with inconsistent data are excluded to for a new time slot collect dt from n devices
go to the next consensus step. The low-rank approximation of 5: d̃t ← project dt on the span of model Eq. (4).
the measurements can be written as follows, 6: z = dt − d̃t .
R
X est for each device (n) that |znS| >model.thresholds(n)
D̃ = λr ur v Tr , (2) 7: Outlier(t) ← Outlier(t) {n}
r=1 end for
where, D̃ is the best low-rank approximation of D. Moreover,
parameter Rest can be estimated by analyzing the singular C. Performance Analysis
values. Rank is the minimum number of learned patterns such A conventional hyperledger fabric network can tolerate up
that data of all devices in a block can be represented as a linear to 1/3 malicious devices. However, proposed AIBC network
combination of those patterns. Using the training data and the can significantly increase this threshold by a carefully de-
estimated rank, a margin around the model can be identified. signed detector. There are two probabilities associated with an
outlier detection algorithm: probability of detection (Pd ) and
TABLE I: Summarization of parameters in the employed model.
probability of false alarm (Pf a ). Probability of miss detection
Model’s (complement of probability of detection) corresponds to the
parameter Variable Description
case that our algorithm fails to detect a malicious node.
Span of the model (regular pat-
terns are linear combination of
Probability of false alarm is associated with the scenario that
Span U ∈ Rb×R
these bases.) the algorithm discerns an intact node as a malicious node.
Estimated The number of basic patterns in According to these two values, the fault tolerance of our
rank Rest the span. proposed architecture can be improved in comparison with the
Threshold for the margin be- naive PBFT consensus protocol. However, there exist the case
Threshold h ∈ Rb tween trustworthy and outlier that the performance of our architecture be less than 33.3%
data.
which is only possible if the detector is not well-designed such
that it has a very high probability of false alarm or a very low
probability of detection.
The impact of the designed detector on the fault tolerance
of the AIBC network is illustrated in the following inequality:
1
Fdet = Fraw (1 − Pd ) + (1 − Fraw )Pf a ≤ . (5)
3
In this inequality, Fdet denotes the fault tolerance of the AIBC
network over the filtered data using the designed detector.
(a) (b)
For successful execution of PBFT consensus in the second
Fig. 4: Acceptable and unavoidable delays. (a) Outlier detection
step consensus, Fdet should be less than 1/3. Fraw is the algorithm delay, (b) Devices state update delay.
initial fault tolerance of our network in the first step before
performing outlier detection which can be greater than 1/3.
However, in a conventional hyperledger fabric, there is no
detector and the fault tolerance (equivalent to Fraw in our
AIBC network) should be less than 1/3. The effect of the
detector on the threshold is investigated via implementation.
IV. I MPLEMENTATION D ETAILS AND R ESULTS
We implemented the proposed 3-layer AIBC network us- (a) (b)
ing hyperledger fabric framework version 1.1.0. The code Fig. 5: Unacceptable and avoidable delays. (a) Model update delay,
is written in chaincode using Golang. Following are the (b) Dataset update delay.
hardware specifications of the two laptops we used: Core i7-
6500U processor, CPU 2.5 GHz × 4, Ubuntu 18.04 LTS. First, discarding it. After a block is sent to the peers by the orderer,
we explain the architecture of the implemented network and invoke function of the detector is executed on each peer for
the dataset we used. Then, we present the performance of execution of step-1 of the consensus protocol. It will discard
the implemented AIBC network in terms of latency and the the outlier data from that block and send the result to the peer
accuracy of the outlier detection algorithm. to initiate the step-2 of the consensus protocol.
A. Network architecture and dataset B. Network Latency
We have simulated the application layer with 100 sensors There are several delays for the invoke function of the
and devices using Matlab. These devices send their data to the designed detector: outlier detection algorithm delay, model
AIBC network in each time unit. The edge blockchain and core update delay, dataset update delay, and devices state update
blockchain layers are simulated on two separate but similar delay. Outlier detection algorithm delay is the time spent to
laptops. Different components of the AIBC network including execute this algorithm to infer outlier data in the latest 100
all the peers and orderer are defined in separate containers readings from 100 devices in the Layer 1. Model update
using Docker. These containers can communicate with each delay is the time it takes to update the model obtained by
other through a channel. The containers in the two laptops are outlier detection algorithm. It should be mentioned that outlier
connected using Docker swarm. detection is a machine learning algorithm in which the learned
The implemented chaincode has three main functions: init, detector is updated at each time based on the observed data.
invoke, and query. Init function is used for initializing the The model is updated based on the last 100 readings from
number of sensors in layer 1 and their names, and initializing different devices in Layer 1. Dataset update delay is the delay
the blockchain with the first input data from the devices. associated with updating the last 100 readings from all 100
Invoke function is used to receive new data from devices devices in the ledger. Devices state update delay is the time
in layer 1. Each time an application sends new data to an spent to update new values of the devices in the ledger.
aggregator, invoke function is executed which sends back the We show the probability density function (pdf) of all the
result after endorsement to that application. Query function is mentioned delays for our implemented IoT network with 100
used to obtain the current value of a device in the IoT network.
1
The detector is also implemented using Golang. It has init 10
and invoke functions. Init function initializes the peers with 100
Delay (sec)

number of devices and their names and initializes primary data

model in AIBC. The primary data model is obtained through 10
-2

outlier detection algorithm from a synthesized dataset. We use

a 100 × 100 matrix where each column is new data obtained 10-4
from an IoT network with 100 devices. Invoke function is
responsible for updating the data model through the outlier 10-6
Dataset Update Model Update Outlier Detection Device Update
detection algorithm, detecting outlier data within a block, and
Fig. 6: Latency of different sections of the AIBC network.
 50
devices in 951 time slots in Figs. 4 and 5. A hyperledger fabric 100

Probability of Detection (%)

57.82%

Probability of Detection (%)

network reach consensus over a new block in milliseconds [6]
45
40
+50%
80
and adding that block to the copy of ledger would take about 35
30
+40% Pfa = 5% and Pd = 46%
5 microseconds according to Fig. 4(b). These delays exist 60
25 Fraw = 57.82%
in any hyperledger fabric implementation and are acceptable 20
+33.3% 40
for a smart home network. Our designed detector would add 15
FAIL ZONE
some new delay components for the network. Outlier detection 10 20
5
algorithm would take about 9.5 milliseconds as shown in 0 0
Fig. 4(a) which will not affect the performance of a smart 5 10 15 20 25 0 20 40 60 80 100
Probability of False Alarm (%) Probability of False Alarm (%)
home network. However, according to Fig. 5, model update (a) (b)
delay and dataset update delay will take about 1.14 and 5.65 Fig. 7: Detector Performance. (a) Fault tolerance of the proposed
seconds respectively which is not acceptable for a smart home algorithm, (b) Accuracy of the proposed outlier detector for a dataset
network. However, dataset update delay and model update with large number of faulty devices.
delay can be easily eliminated if we use pre-learned data using
however, recent advances in artificial intelligence such as
a proper dataset. Therefore, our proposed architecture will
deep learning and reinforcement learning can be exploited for
incur an additional delay of about 9.5 milliseconds for outlier
designing the detector in a more robust manner.
detection which is acceptable for a smart home network since
consensus is reached in milliseconds. The delays for different R EFERENCES
sections of our implementation are compared in Fig. 6.
[1] T.M. Fernndez-Carams and P. Fraga-Lamas. ”A Review on the Use of
C. Two-step Consensus Protocol Accuracy Blockchain for the Internet of Things.” IEEE Access 6 (2018).
[2] J. Lin, Z. Shen, and C. Miao. ”Using blockchain technology to build
The last factor to evaluate our implementation is the ac- trust in sharing LoRaWAN IoT.” In Proceedings of the 2nd International
curacy of our proposed algorithm in terms of probability of Conference on Crowd Science and Engineering, pp. 38-43. ACM, 2017.
[3] O. Novo. ”Blockchain meets IoT: An architecture for scalable access
detection and probability of false alarm and its impact on management in IoT.” IEEE Internet of Things Journal 5, no. 2 (2018).
the fault tolerance of the proposed architecture. The fault [4] Y. Dai, D. Xu, S. Maharjan, Z. Chen, Q. He and Y. Zhang. (2019).
tolerance for different probabilities of detection and false alarm Blockchain and deep reinforcement learning empowered intelligent 5g
beyond. IEEE Network, 33(3), 10-17.
is illustrated in Fig. 7(a). Fault tolerance of less than 33.3% is [5] M. Salimitari, M. Chatterjee, M. Yuksel, and E. Pasiliao. ”Profit max-
denoted as fail zone which is only possible if the detector is imization for bitcoin pool mining: A prospect theoretic approach.” In
designed very inexpertly with a significantly high probability 2017 IEEE 3rd International Conference on Collaboration and Internet
Computing (CIC), pp. 267-274. IEEE, 2017.
of false alarm or low probability of detection. In general, [6] M. Salimitari, and M. Chatterjee. ”A survey on consensus protocols
the performance of the network is enhanced and the fault in blockchain for IoT networks.” arXiv preprint arXiv:1809.05613v4
tolerance of the AIBC network for different detectors (different (2018).
[7] A. Reyna, C. Martn, J. Chen, E. Soler, and M. Daz. ”On blockchain
probability of detection and probability of false alarm) is found and its integration with IoT. Challenges and opportunities.” Future
to be more than 33.3%, more than 40%, or more than 50% as Generation Computer Systems 88 (2018): 173-190.
shown in Fig. 7(a). [8] Z. Zheng, S. Xie, H. Dai and H. Wang. ”Blockchain challenges and
opportunities: A survey.” Work Pap.2016 (2016).
Fig. 7(b) shows the accuracy of our algorithm on a synthe- [9] Z. Zheng, S. Xie, H. Dai, X. Chen and H. Wang. ”An overview of
sized dataset with large number of faulty devices. Although blockchain technology: Architecture, consensus, and future trends.” In
there exists a large number of malicious devices (outlier 2017 IEEE International Congress on Big Data (BigData Congress), pp.
557-564. IEEE, 2017.
data) in the dataset, the detector is learned such that the [10] S. Gupta and M. Sadoghi. ”Blockchain Transaction Processing.” (2019):
fault tolerance of the network is more than 50%. This is 1-11.
inferred by choosing any operating point in Fig. 7(b) and the [11] T. N. Dinh and M. T. Thai. ”Ai and blockchain: A disruptive integration.”
Computer 51, no. 9 (2018): 48-53.
corresponding point in Fig. 7(a). As an example, an operating [12] S. Dey. ”Securing majority-attack in blockchain using machine learning
point of the detector is shown at Pf a = 5% and Pd = 46%. and algorithmic game theory: A proof of work.” In 2018 10th Computer
The fault tolerance of the network in this point is 57.82% Science and Electronic Engineering (CEEC), pp. 7-10. IEEE, 2018.
[13] M. Udell, C. Horn, R. Zadeh and S. Boyd. ”Generalized low rank
according to Inequality (5). models.” Foundations and Trends in Machine Learning 9, no. 1 (2016).
[14] T. Baltruaitis, C. Ahuja and L.-P. Morency. ”Multimodal machine learn-
V. C ONCLUSIONS ing: A survey and taxonomy.” IEEE Transactions on Pattern Analysis
and Machine Intelligence 41, no. 2 (2018): 423-443.
We proposed the AIBC network with a 2-step consensus [15] M. Joneidi, P. Ahmadi, M. Sadeghi and N. Rahnavard. ”Union of low-
protocol using outlier detection algorithm and PBFT. Outlier rank subspaces detector.” IET Signal Processing 10, no. 1 (2016): 55-62.
[16] A. Esmaeili and F. Marvasti. ”A novel approach to quantized matrix
detection algorithm acts as the first step consensus and verifies completion using huber loss measure.” IEEE Signal Processing Letters
the compatibility of new data and discards the suspicious 26, no. 2 (2019): 337-341.
ones in order to increase fault tolerance of the network for [17] S. Li, M. Shao and Y. Fu. ”Multi-view low-rank analysis with applica-
tions to outlier detection.” ACM Transactions on Knowledge Discovery
the second step consensus (PBFT). We measured the latency, from Data (TKDD) 12, no. 3 (2018): 32.
accuracy, and performance of our method. Results reveal [18] H. Zhao, H. Liu, Z. Ding and Y. Fu. ”Consensus regularized multi-view
significant increase in the fault tolerance of hyperledger fabric outlier detection.” IEEE Transactions on Image Processing 27, no. 1
(2017): 236-248.
by our detector. We employed an outlier detection scheme,