Adversarial Robustness of Spiking Neural Networks

Determine the adversarial robustness of Spiking Neural Networks (SNNs) used in embedded and security-critical systems.

Background

The paper investigates adversarial example attacks on Spiking Neural Networks (SNNs), which are touted for energy efficiency and biological plausibility, and are increasingly used in embedded and security-critical applications. Despite these advantages, the extent of SNNs’ resistance to adversarial perturbations has not been conclusively established.

To address this gap, the authors propose Spike-PTSD, a biologically inspired attack method that exploits neuron hyper/hypoactivation patterns analogous to those observed in PTSD-affected brains. While the method demonstrates high empirical attack success rates across datasets and models, the broader question of SNN adversarial robustness remains unresolved as explicitly stated in the abstract.