Official implementation for paper Generating Minimalist Adversarial Perturbations to Test Object-Detection Models: An Adaptive Multi-Metric Evolutionary Search Approach.

Deep Learning (DL) models excel in computer vision tasks but can be susceptible to adversarial examples. This paper introduces Triple-Metric EvoAttack (TM-EVO), an efficient algorithm for evaluating the robustness of object-detection DL models against adversarial attacks. TM-EVO utilizes a multi-metric fitness function to guide an evolutionary search efficiently in creating effective adversarial test inputs with minimal perturbations. We evaluate TM-EVO using widely-used object-detection DL models, DETR and Faster R-CNN, and open-source datasets, COCO and KITTI. Our findings reveal that TM-EVO outperforms the state-of-the-art EvoAttack baseline, leading to better adversarial tests. TM-EVO's balance of effectiveness and efficiency represents an advancement in adversarial test generation.

Before installing the dependencies, it is recommended to create a virtual environement. If you are using conda, the following commands can be executed to create and activate the virtual environment.

conda create --name TM-EVO python=3.9.17
conda activate TM-EVO

The required dependencies have been frozen in a requirements.txt file. Once the virtual environment has been created and activated, the following commande can be executed to install the dependencies.

pip install -r requirements.txt

We use the open-source high-quality dataset tool Fiftyone to download and load the KITTI and COCO datasets. A dataset is downloaded the first time an experiment which uses said dataset is executed. Once downloaded, subsequent executions of experiments will load the dataset.

Note - The entire dataset will be installed even though only a subset of images will be used during experiments.

We use object-detection models DETR and Faster R-CNN under test. We use the Huggingface implementation of DETR and the Pytorch implementation of Faster R-CNN. A pretrained model is downloaded the first time an experiment which uses said model is executed. Once downloaded, subsequent executions of experiments will load the model.

We provide experiments.py with which experiments can run on a combination of algorithms, datasets and models.

• data_location is a required argument. it is the file path to the root folder where the datasets will be saved.
• algs is the algorithm that will be tested during the experiment. Available commands are EVO, TM_EVO or ALL.
• datasets is the dataset that will be tested during the experiment. Available commands are COCO, KITTI or ALL.
• mdels is the model that will be tested during the experiment. Available commands are DETR, FASTER or ALL.
• dataset_size is the number of images to process during the experiment. The default value is 10.
• runs is the number of runs per image to process during the experiment. The default value is 5.

The following arguments are available and will change the behaviour of the genetic algorithm.

• population is the population size. The default value is 32.
• generations is the max number of generations. The default value is 600.
• p_mut is the mutation probability. The default value is 0.012.
• p_unmut is the is the mutation reduction probability. The default value is 0.3.
• delta is the max degree of perturbation of a pixel when mutated. The default value is 0.4.

For example, the following command will run an experiments on both the DETR and FASTER R-CNN models, 5 images from the COCO dataset, will use the TM_EVO algorithm to perform the adversarial input generation, and only 1 run of the experiment will be performed on each image.

python experiment.py /path/to/dataset/root --algs TM_EVO --datasets COCO --dataset_size 5 --runs 1

A logs folder will be created automatically where information about the experiment will be saved using the following structure.

/logs
.    log_1/
.    .    parameters.txt
.    .    summary.txt
.    .    model/
.    .    .    adaptive/
.    .    .    .    dataset/
.    .    .    .    .    algorithm/
.    .    .    .    .    .    all_fitness_1.txt
.    .    .    .    .    .    all_norm_0_1.txt
.    .    .    .    .    .    all_norm_2_1.txt
.    .    .    .    .    .    fitness_1.txt
.    .    .    .    .    .    generations_1.txt
.    .    .    .    .    .    norm_0_1.txt
.    .    .    .    .    .    norm_2_1.txt
.    .    .    .    .    .    run_time_1.txt
.    .    .    .    .    .    ...
.    .    .    .    .    .    all_fitness_j.txt
.    .    .    .    .    .    all_norm_0_j.txt
.    .    .    .    .    .    all_norm_2_j.txt
.    .    .    .    .    .    fitness_j.txt
.    .    .    .    .    .    generations_j.txt
.    .    .    .    .    .    norm_0_j.txt
.    .    .    .    .    .    norm_2_j.txt
.    .    .    .    .    .    run_time_j.txt

    ...

.    log_i]

A imgs folder will be created automatically where information about the experiment will be saved using the following structure.

/imgs/
.    AEs/
.    .    [1]/
.    .    .    model/
.    .    .    .    dataset/
.    .    .    .    .    algorithm/
.    .    .    .    .    .    imgs1.png
.    .    .    .    .    .    ...
.    .    .    .    .    .    imgsk.png

        ...

.    .    i/

In the above file structure, i is the index of the current experiment,j is the number of runs of the experiment per image, and k is the number of images generated (number of images sampled from the experiment multiplied by the number of runs).

The files can be described as follows.

• parameters.txt is a file containing all the parameter information of the experiment.
• summary.txt is a summary of all the results of the experiment, including norm information and execution time.
• all_fitness_j.txt is a file containing all fitness history of the runs.
• all_norm_0_j.txt is a file containing all L0 norm history of the runs.
• all_norm_2_j.txt is a file containing all L2 norm history of the runs.
• fitness_j.txt is the final fitness of the runs (0 if successful).
• generations_j.txt is the number of generations passed by the algorithm.
• norm_0_j.txt is the final L0 norm of the runs.
• norm_2_j.txt is the final L2 norm of the runs.
• run_time_j.txt is the execution time of the runs.
• imgsk.png.txt is an images that can be used as an adversarial input.

In section we compare adversarial test inputs generated by the baseline EvoAttack algorithm and our TM-EVO alogrithm on the KITTI and COCO datasets.

(a) (b)

(c) (d)

Figure 1: Sucessfull adversarial test inputs for the KITTI dataset (a) Attack performed on the Faster R-CNN model with EvoAttack baseline (b) Attack performed on the Faster R-CNN model with our algorithm TM-EvoAttack (c) Attack performed on the Detr model with EvoAttack baseline (d) Attack performed on the Detr model with our algorithm TM-EvoAttack

(a) (b)

(c) (d)

Figure 2: Sucessfull adversarial test inputs for the COCO dataset (a) Attack performed on the Faster R-CNN model with EvoAttack baseline (b) Attack performed on the Faster R-CNN model with our algorithm TM-EvoAttack (c) Attack performed on the Detr model with EvoAttack baseline (d) Attack performed on the Detr model with our algorithm TM-EvoAttack.

All images can be found in the report folder.

In this section we show the quantitative results of our experiment.

Figure 3: Empirical evaluation on the datasets, models and algorithms.

All the data, along with the visuals, can be found in the data report folder.

This repository is released under the Apache 2.0 license as found in the LICENSE file.