Time-Distributed Backdoor Attacks on Federated Spiking Learning

Gorka Abad, Stjepan Picek, Aitor Urbieta

Abstract: This paper investigates the vulnerability of federated learning (FL) with spiking neural networks (SNNs) to backdoor attacks using neuromorphic data. Despite the efficiency of SNNs and the privacy advantages of FL, particularly in low-powered devices, we demonstrate that these systems are susceptible to such attacks. We first assess the viability of using FL with SNNs using neuromorphic data, showing its potential usage. Then, we evaluate the transferability of known FL attack methods to SNNs, finding that these lead to sub-optimal attack performance. Consequently, we explore backdoor attacks involving single and multiple attackers to improve the attack performance. Our main contribution is developing a novel attack strategy tailored to SNNs and FL, which distributes the backdoor trigger temporally and across malicious clients, enhancing the attack’s effectiveness and stealthiness. In the best case, we achieve a 100% attack success rate, 0.13 MSE, and 98.9 SSIM. Moreover, we adapt and evaluate existing defenses against backdoor attacks, revealing their inadequacy in protecting SNNs. Our code is publicly available. (https://github.com/GorkaAbad/Time-Bandits).

External IDs:dblp:conf/esorics/AbadPU25