Introduction
Spiking Neural Networks (SNNs) represent a biologically inspired, energy-efficient alternative to traditional Artificial Neural Networks (ANNs). By communicating via discrete spikes rather than continuous values, SNNs are ideally suited for deployment in edge computing and autonomous systems. However, their non-differentiable nature necessitates the use of Surrogate Gradients (SG) during training. While SGs enable backpropagation, they also introduce a significant security vulnerability: SNNs become highly susceptible to adversarial attacks, where imperceptible input perturbations can lead to catastrophic misclassification.
Addressing the Robustness Gap
The vulnerability of SNNs is fundamentally linked to the magnitude of their gradients. High-sensitivity regions in the network's internal state space allow small input changes to propagate and amplify, leading to incorrect outputs. This research introduces MPD-SGR (Membrane Potential Distribution-driven Surrogate Gradient Regularization), a novel technique designed to enhance the structural robustness of SNNs by managing the distribution of internal neuronal states.
Methodology: MPD-SGR
The core of MPD-SGR lies in aligning the statistical distribution of a neuron's membrane potential ($U$) with the active window of the surrogate gradient function.
The researchers derived a theoretical link between robustness error and gradient magnitude. By modeling membrane potentials as a Gaussian distribution—a characteristic observed in SNNs using threshold-dependent Batch Normalization—they identified the "overlap area" ($\Omega$) between this distribution and the SG's active region. MPD-SGR introduces a regularization term that minimizes this overlap, effectively pushing more neurons into a stable zone where they are less reactive to noise.
Experimental Evaluation
The proposed method was evaluated on CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets using VGG11 and WideResNet (WRN16) architectures. The models were subjected to various adversarial attacks, including FGSM, PGD, BIM, and CW.
Key Results
MPD-SGR demonstrates substantial improvements in adversarial robustness across all tested scenarios.
- Significant Robustness Gains: On CIFAR-10 under strong PGD attacks, MPD-SGR improved accuracy from near 0% (vanilla) to 20.55%, more than doubling the performance of previous state-of-the-art (SOTA) methods.
- Synergy with Adversarial Training: Combining MPD-SGR with Adversarial Training (AT) further boosted performance, reaching 33.38% accuracy on CIFAR-10 under PGD, outperforming the FEEL method by nearly 4%.
- Structural Resilience: The method provides genuine resilience rather than mere "gradient masking," maintaining effectiveness across different surrogate functions and spike encoding schemes.
Discussion and Future Directions
While MPD-SGR significantly enhances robustness, it involves a minor trade-off in "clean" accuracy (typically a 1-2% decrease). Additionally, the regularization strength hyperparameter ($\eta$) requires careful tuning to balance learning capacity with robustness. Future research will focus on scaling these SNN-specific regularizers to larger datasets like ImageNet-1K and further optimizing the trade-off between efficiency and security.
Summary updated