Artificial VO2 Spiking Neurons with Protective Mechanism for Enhancing Resilience of Spiking Neural Network Against Adversarial Attacks

Spiking Neural Networks (SNNs) have demonstrated significant potential in power-efficient edge computing. However, SNNs are vulnerable to adversarial attacks which seriously affects the accuracy and thus limits its application. In this study, a defense mechanism against adversarial attacks is demonstrated based on VO
2
spiking neuron. The spiking neuron circuit (SNC) can mimic the protective mechanism of neurons to effectively filter adversarial perturbations among input signals. Furthermore, we evaluate the ability of SNC-based SNNs against adversarial attacks with MNIST task. The SNC-based SNNs can effectively withstand adversarial attacks, achieving an 18.87% improvement in accuracy compared to the ReLU method under 30% disturbance. The results demonstrate that the SNC method has significant advantage in enhancing the SNN system's ability to resist adversarial attacks. This research provides an innovative solution to tackle the security issue in highly reliable neuromorphic computing.