Reproduce “Towards Reliable Evaluation of Adversarial Robustness for Spiking Neural Networks” on OpenResearch

About this paper

Jihang Wang, Dongcheng Zhao, Ruolin Chen, Qian Zhang, Yi Zeng

A novel framework combining an Adaptive Sharpness Surrogate Gradient (ASSG) method and a Stable Adaptive Projected Gradient Descent (SA-PGD) algorithm provides a more reliable assessment of adversarial robustness in Spiking Neural Networks (SNNs). This approach reveals that SNNs are significantly more vulnerable to adversarial attacks than prior evaluations suggested.

Key insights

• The degree of gradient vanishing in SNN surrogate functions can be precisely controlled by adaptively adjusting sharpness, significantly improving the accuracy of gradient estimation for adversarial attacks.
• Effective adversarial attack optimization against SNNs requires not only improved gradient approximations but also robust algorithms like SA-PGD that can handle imprecise gradients and ensure stable convergence.
• Existing SNN adversarial robustness evaluations are often optimistic, as current adversarial training methods using basic surrogate gradients and few attack steps are insufficient to uncover the true vulnerabilities of SNNs.

Summary by alphaXiv.