SNN-GST: Gradient-Based Security Testing Method for Spiking Neural Networks

Spiking Neural Networks (SNNs) have garnered significant attention in the field of deep learning due to their efficient energy utilization and inspiration drawn from neuroscience. However, the security issues of SNNs, particularly their vulnerability in safety-critical tasks, have attracted the attention of researchers. Therefore, we have investigated the security issues of SNNs and designed a security testing method for SNNs, termed SNN-GST. This method leverages an understanding of the working principles and vulnerabilities of SNNs to conduct gradient-based adversarial attacks. It employs gradient optimization techniques to find solutions and adjusts the spiking threshold to reduce ineffective spikes. Additionally, it introduces a perturbation norm constraint method to decrease the average perturbation required for the attack. Furthermore, we have designed a method and metrics for evaluating the temporal robustness of SNNs. By integrating the SNN-GST method for conducting attack tests, we have assessed the robustness of the models across various datasets. This study aims to design the SNN-GST for a comprehensive quantitative assessment of the robustness of SNNs, providing scientific methods and metrics to enhance their security, and ensuring the reliability of SNNs in practical applications.