Target-X

This repository introduces Target-X, a novel and fast method for the construction of adversarial targeted images on large-scale datasets that can fool the state-of-the-art image classification neural networks. We evaluate the performance of Target-X using the well-trained image classification neural networks of different architectures and compare it with the well-known T-FGSM and T-UAP targeted attacks. The reported results demonstrate that Target-X can generate targeted adversarial images with the least perturbations on large-scale datasets that can fool the image classification neural networks and significantly outperform the T-FGSM and T-UAP attacks.

Results

Here we show our work, Target-X, compared to two other adversarial attacks, T-FGSM and T-UAP. We compare with four metrics:

A_f, denoting the average prediction accuracy of the architecture on the original images.
P_f, denoting the average robustness of the perturbations on the the prediction accuracy of the architecture.
T_f, denoting the success rate of targeted perturbations.
E_f, denoting the average execution time.

Four tables are given, each showing the result of testing with different perturbation sizes (Ɛ). Areas where Target-X outperforms other techniques are shown in bold.

Ɛ = 0.0005

Algorithm	Architecture	A_f	P_f	T_f	E_f
Target-X	AlexNet	0.573	0.302	0.596	0.7754
	DenseNet-201	0.766	0.184	0.803	3.9975
	GoogleNet	0.697	0.129	0.859	0.8496
	ResNet-34	0.739	0.108	0.867	2.4447
	Resnet-101	0.775	0.155	0.824	0.8557
	VGG-19	0.730	0.292	0.618	2.7172
T-FGSM	AlexNet	0.573	0.568	0.0380	0.0187
	DenseNet-201	0.766	0.762	0.0418	0.1566
	GoogleNet	0.697	0.694	0.0520	0.0466
	ResNet-34	0.739	0.730	0.0528	0.0840
	ResNet-101	0.775	0.772	0.0492	0.0350
	VGG-19	0.730	0.725	0.0550	0.0723
T-UAP	AlexNet	0.573	0.573	0.0110	0.4741
	DenseNet-201	0.766	0.766	0.0150	3.9476
	GoogleNet	0.697	0.697	0.0120	1.1437
	ResNet-34	0.739	0.739	0.0080	2.0902
	ResNet-101	0.775	0.775	0.0012	0.8952
	VGG-19	0.730	0.730	0.0080	1.7886

Ɛ = 0.005

Algorithm	Architecture	A_f	P_f	T_f	E_f
Target-X	AlexNet	0.573	0.066	0.941	0.2242
	DenseNet-201	0.766	0.029	0.991	0.8817
	GoogleNet	0.697	0.036	0.986	0.1952
	ResNet-34	0.739	0.026	0.992	0.5791
	Resnet-101	0.775	0.032	0.983	0.1760
	VGG-19	0.730	0.080	0.887	1.0760
T-FGSM	AlexNet	0.573	0.464	0.4048	0.0184
	DenseNet-201	0.766	0.547	0.4332	0.1568
	GoogleNet	0.697	0.476	0.4818	0.0461
	ResNet-34	0.739	0.461	0.5094	0.0844
	ResNet-101	0.775	0.535	0.4498	0.0348
	VGG-19	0.730	0.449	0.5300	0.0734
T-UAP	AlexNet	0.573	0.573	0.0090	0.4629
	DenseNet-201	0.766	0.766	0.0100	3.9603
	GoogleNet	0.697	0.697	0.0100	1.1440
	ResNet-34	0.739	0.739	0.0120	2.0899
	ResNet-101	0.775	0.775	0.0130	0.8991
	VGG-19	0.730	0.730	0.0100	1.7888

Ɛ = 0.05

Algorithm	Architecture	A_f	P_f	T_f	E_f
Target-X	AlexNet	0.573	0.038	0.980	0.0831
	DenseNet-201	0.766	0.022	0.997	0.3809
	GoogleNet	0.697	0.030	0.996	0.0989
	ResNet-34	0.739	0.022	0.996	0.2766
	Resnet-101	0.775	0.022	0.993	0.0883
	VGG-19	0.730	0.054	0.923	0.5926
T-FGSM	AlexNet	0.573	0.062	0.9620	0.0186
	DenseNet-201	0.766	0.122	0.9010	0.1565
	GoogleNet	0.697	0.107	0.9164	0.0458
	ResNet-34	0.739	0.077	0.9410	0.0838
	ResNet-101	0.775	0.161	0.8590	0.0349
	VGG-19	0.730	0.075	0.9456	0.0701
T-UAP	AlexNet	0.573	0.573	0.0110	0.4631
	DenseNet-201	0.766	0.766	0.0100	3.9629
	GoogleNet	0.697	0.697	0.0090	1.1443
	ResNet-34	0.739	0.739	0.0090	2.0854
	ResNet-101	0.775	0.775	0.0090	0.9002
	VGG-19	0.730	0.731	0.0120	1.7865

Ɛ = 0.2

Algorithm	Architecture	A_f	P_f	T_f	E_f
Target-X	AlexNet	0.573	0.034	0.981	0.0826
	DenseNet-201	0.766	0.024	0.999	0.3818
	GoogleNet	0.697	0.027	0.996	0.1059
	ResNet-34	0.739	0.027	0.997	0.3043
	Resnet-101	0.775	0.024	0.993	0.0917
	VGG-19	0.730	0.041	0.949	0.5616
T-FGSM	AlexNet	0.573	0.024	0.9838	0.0184
	DenseNet-201	0.766	0.145	0.8740	0.1569
	GoogleNet	0.697	0.131	0.8864	0.0458
	ResNet-34	0.739	0.093	0.9224	0.0840
	ResNet-101	0.775	0.185	0.8262	0.0351
	VGG-19	0.730	0.067	0.9474	0.0702
T-UAP	AlexNet	0.573	0.572	0.0100	0.4628
	DenseNet-201	0.766	0.765	0.0120	3.9903
	GoogleNet	0.697	0.697	0.0110	1.1463
	ResNet-34	0.739	0.739	0.0090	2.0850
	ResNet-101	0.775	0.775	0.0090	0.9007
	VGG-19	0.730	0.730	0.0090	1.7874

Setup and Installation

Cloning from git

To clone the repository us the following:

git clone https://github.com/nssrlab/target-x.git <DESTINATION>

Where <DESTINATION> is the folder you want Target-X to reside in.

Setting up a Conda environment

To create a conda environment to run the project in, use the provided environment.yml in the following command:

conda env create -f environment.yml -n <ENVIRONMENT_NAME>

Where <ENVIRONMENT_NAME> is the name you want to use for the conda environment.

Cite Us

@INPROCEEDINGS{10197071,
  author={Khamaiseh, Samer Y. and Bagagem, Derek and Al-Alaj, Abdullah and Mancino, Mathew and Alomari, Hakem and Aleroud, Ahmed},
  booktitle={2023 IEEE 47th Annual Computers, Software, and Applications Conference (COMPSAC)}, 
  title={Target-X: An Efficient Algorithm for Generating Targeted Adversarial Images to Fool Neural Networks}, 
  year={2023},
  volume={},
  number={},
  pages={617-626},
  keywords={Training;Image recognition;Perturbation methods;Force;Artificial neural networks;Computer architecture;Robustness;adversarial images;deep learning;image classification neural networks;adversarial deep neural networks},
  doi={10.1109/COMPSAC57700.2023.00087}}

LAiSR-SK/target-x