yudidong / gan-based-e2e-communications-system-for-defense-against-adversarial-attack

A Robust Adversarial Network-Based End-to-End Communications System With Strong Generalization Ability Against Adversarial Attacks

This is an implement of our pre-print paper:

Yudi Dong and Huaxia Wang and Yu-Dong Yao, “A Robust Adversarial Network-Based End-to-End Communications System With Strong Generalization Ability Against Adversarial Attacks”, https://arxiv.org/abs/2103.02654

This paper has been accepted by IEEE ICC 2022.

TensorFlow Version

Our codes are based on TensorFlow-GPU 2.0

Main Function Files:

"gan_blackbox.py": BLER Peformance of our proposed method under black-box attacks

"gan_whitebox.py": BLER Peformance of our proposed method under white-box attacks

"regular_training_blackbox.py": BLER Peformance of regular training method under black-box attacks

"regular_training_whitebox.py": BLER Peformance of regular training method under white-box attacks

"adversarial_training_blackbox.py": BLER Peformance of adversarial training method under black-box attacks

"adversarial_training_whitebox.py": BLER Peformance of adversarial training method under white-box attacks

Class Function Files:

"classes/GAN_Classes.py": Implement for our proposed GAN-based end-to-end system

"classes/Autoencoder_Classes.py": Implement for the autoencoder end-to-end system

"classeshamming.py": Implement for the traditional communications system (BPSK, Hamming)

Other Files:

"UAP": perturbations used for black-box attacks

Coding Reference:

[1] https://github.com/meysamsadeghi/Security-and-Robustness-of-Deep-Learning-in-Wireless-Communication-Systems

gan-based-e2e-communications-system-for-defense-against-adversarial-attack's Issues

Scale_factor calculate and Variable 'n'

Hello, I have a question about the variable 'n' . In your paper, you describe n si ''n refers to the number of channel uses''.So in your experiment, the shape of x and UAP is (1,2,n), right? But, when calculate the eps(scale_factor = np.sqrt( (PSR * self.n) / (np.linalg.norm(p)**2 + 0.00000001)) ) , if the length of channels used on UAP and x is different,while length can I used to calculate the scale_factor ?
Thank you so much!