A Bio-Inspired Hierarchical Temporal Defense for Securing Spiking Neural Networks Against Physical and Adversarial Perturbations

Spiking Neural Networks (SNNs) offer an energy-efficient paradigm for processing temporal data but are critically vulnerable to perturbations in spike timing, a common issue in physical systems where jitter arises from sensor noise or radiation-induced effects. Traditional defenses adapted from static neural networks often fail to address these unique temporal dynamics. We introduce the Hierarchical Temporal Defense (HTD), a novel, bio-inspired architecture integrating defenses across the input, neuronal, and synaptic levels. Key innovations include probabilistic encoding for jitter tolerance, adaptive thresholds for stability, and gated plasticity for secure learning. Theoretical analysis provides formal robustness guarantees, including finite KL-divergence bounds under jitter. Empirical evaluation on high-fidelity simulated physical data shows the HTD framework reduces the success rate of a strong PGD attack from 82.1\% to 18.7\% and maintains high performance (F1 > 0.72) under environmental stress, demonstrating a principled methodology for designing robust neuromorphic systems for safety-critical applications.