BitFool: An Optimization-Based Adversarial Attack for End-to-End Communication Systems

End-to-end communication is an efficient paradigm that combines autoencoder neural networks (NN) with traditional communication systems. However, it inherits the vulnerability to adversarial attacks from deep learning. In this paper, we consider a bit-wise end-to-end communication system under malicious attack, where the attacker introduces physical perturbations to degrade the system's bit error rate (BER) performance. We analyze several common adversarial perturbations and then we propose a framework for crafting gradient-based attacks. Besides, based on the channel layer visualization, we analyze the model's learned signal representation and propose an upper bound of the attack performance. Meantime, inspired by deep learning (DL), we design an optimization-based adversarial attack called BitFool which is more destructive and computationally efficient. The simulation results in terms of BER show that BitFool is superior to other methods and performs very close to the upper bound.