Random heterogeneous spiking neural network for adversarial defense

Spiking neural networks (SNNs) offer a biologically inspired alternative to artificial neural networks (ANNs) by mimicking neuronal information transmission mechanisms. However, similar to ANNs, SNNs remain susceptible to adversarial examples, raising concerns about their robustness in practical applications. To address this vulnerability, we propose the Random Heterogeneous Spiking Neural Network (RandHet-SNN), inspired by the heterogeneity and stochasticity observed in biological neural systems. This architecture strengthens the network's defense against adversarial attacks by introducing neuron-level diversity through randomized time decay constants, allowing each neuron to acquire unique temporal properties at every forward pass. We evaluate RandHet-SNN's performance through extensive experiments with various adversarial attacks. Results indicate that RandHet-SNN significantly enhances network robustness while maintaining minimal impact on clean accuracy. RandHet-SNN displays significant potential for robust, energy-efficient neural computing in adversarial environments.