Momentum-Based Adversarial Attacks Against End-to-End Communication Systems

Deep learning (DL) based communication system is a promising novel architecture to implement end-to-end optimization compared with conventional block-separated optimization schemes. However, the vulnerability to adversarial examples of deep neural networks poses significant security concern on the end-to-end communication systems. Adversarial attacks serve as a fundamental surrogate to evaluate the robustness of the DL-based communication systems before they are deployed. Specifically, we propose a new adversarial attack method with momentum iterative gradient against the end-to-end communication systems. For targeted attacks, embedding the momentum term in the iterative process can help loss function stabilize the update direction and avoid getting stuck in saddle points and poor local minima. Therefore, the momentum-based method can enhance the effectiveness without losing the transferability of adversarial attacks. Numerous simulation results illustrate that the proposed method can achieve superior block error rate compared with traditional jamming attacks and no momentum accumulated adversarial attacks.