Spiking Neural Networks leverage event-driven processing and spatiotemporal sparsity for low-power neuromorphic computing, yet they are vulnerable to input perturbations, hardware noise, and adversarial attacks in real deployments. Conventional SNNs trained with cross‑entropy tend to encode input‑specific hidden representations; an information‑plane analysis shows that substantial input information persists even after label‑predictive information stabilizes.
To mitigate this, we introduce SCIB (Conditional Information Bottleneck). Unlike traditional variational IB, SCIB replaces the compression term $I(X;S)$ with the conditional mutual information $I(X;S|Y)$ and derives a tractable variational upper bound. It also employs a learnable class‑conditional prior instead of a label‑agnostic isotropic Gaussian, enhancing the expressiveness of compression and encouraging class‑structured representations when combined with the supervised loss.
We evaluate SCIB on Fashion‑MNIST, CIFAR‑10, and CIFAR‑100 under external environmental noise, neuromorphic hardware noise, and several adversarial attacks. Results indicate that SCIB improves robustness over standard SNNs and IB‑based baselines across most noise settings, raises adversarial accuracy without adversarial training, and reduces SynOps‑based single‑sample inference energy by roughly 24.9%.
In summary, the conditional information bottleneck offers a simple yet powerful mechanism to constrain SNN hidden states toward label‑relevant information, thereby enhancing robustness and energy efficiency in noisy and adversarial scenarios.
Review