Research map / Adversarial robustness and security / Attacks on text and graphs
Enhancing the Robustness of Spiking Neural Networks with Stochastic Gating Mechanisms
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a stochastic gating (StoG) model that Bernoulli-samples spike transmissions between SNN layers, with theory linking it to a regularizer, Lipschitz reduction, and Poisson coding, improving adversarial robustness on CIFAR-10/100.
Problem
SNNs remain vulnerable to adversarial attacks and noise perturbations, and the role of stochasticity in improving their robustness is poorly understood.
Approach
Insert a learnable stochastic gate, sampled from a Bernoulli distribution with per-neuron opening probability, between layers to filter spikes; theory shows this bounds adversarial loss change and layer-wise error amplification and explains Poisson coding as a special case.
Key results
- On CIFAR-10 VGG-11, RAT+StoG raises PGD-l∞ untargeted robust accuracy from 0.04% (vanilla) and 23.37% (AT) to 27.74% (AT+StoG) and 39.58% (RAT+StoG) (RAT+StoG ~ 39.58% vs RAT 28.08%).
- Under EOT-AutoPGD on CIFAR-10 VGG-11, Vanilla+StoG improves robust accuracy from 2.11% to 11.58%.
- A larger gate-penalty gamma=1e-5 reduces average opening probability to 0.52 and SOP by 1.428x for VGG-11 while only slightly lowering clean accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add stochastic gating to RAT | VGG-11 · CIFAR-10, FGSM-UT | robust accuracy: 48.46 -> 55.98 | Helped |
| Add stochastic gating to RAT | VGG-11 · CIFAR-10, PGD-l∞-UT | robust accuracy: 28.08 -> 39.58 | Helped |
| Add stochastic gating to RAT | VGG-11 · CIFAR-100, FGSM-UT | robust accuracy: 29.23 -> 33.40 | Helped |
| Add stochastic gating to RAT | VGG-11 · CIFAR-100, PGD-l∞-UT | ASR: 75.49 -> 62.99 | Helped |
| Add stochastic gating to vanilla training | VGG-11 · CIFAR-10, clean | accuracy: 93.05 -> 91.64 | Hurt |
| Add stochastic gating to vanilla training | VGG-11 · CIFAR-100, clean | accuracy: 73.38 -> 70.44 | Hurt |
| Increase the StoG penalty strength γ (ablation) | WRN-16-4 · CIFAR-10 | clean accuracy | No clear effect |
| Increase the StoG penalty strength γ (ablation) | VGG-11 · CIFAR-10 | spike operations (SOP): 1.428× | Helped |
| Increase the StoG penalty strength γ (ablation) | WRN-16-4 · CIFAR-10, FGSM (ϵ = 2/255) | robust accuracy | Helped |
Surprising findings
- Adding StoG to the vanilla model cut clean accuracy only modestly while improving robustness, and the paper reports no clear gradient obfuscation under randomly initialized EOT-PGD.
Limitations
- Low gate opening probabilities filter out most signals, sacrificing some clean accuracy.
- SNN-specific gradient attacks are not well studied.
- Robustness evaluated only with limited EOT steps; larger EOT steps should be investigated.
Experimental setup
Datasets: CIFAR-10 CIFAR-100
Benchmarks: CIFAR-10 CIFAR-100
Metrics: robust accuracy attack success rate spike operations (SOP)
Baselines: Vanilla BPTT SNN adversarial training (AT) RAT (AT with Lipschitz penalty)
Builds on: Poisson coding BPTT with surrogate gradients FGSM adversarial training Lipschitz regularized SNN training (Ding et al. 2022) Gumbel-Max reparameterization
Closest papers by meaning
- Enhancing Training of Spiking Neural Network with Stochastic Latency (AAAI 2024 · similarity 0.90)
- Robust Stable Spiking Neural Networks (ICML 2024 · similarity 0.89)
- FEEL-SNN: Robust Spiking Neural Networks with Frequency Encoding and Evolutionary Leak Factor (NeurIPS 2024 · similarity 0.89)
- Enhancing Adversarial Robustness in SNNs with Sparse Gradients (ICML 2024 · similarity 0.89)
- Autaptic Synaptic Circuit Enhances Spatio-temporal Predictive Learning of Spiking Neural Networks (ICML 2024 · similarity 0.87)
- Rethinking the Membrane Dynamics and Optimization Objectives of Spiking Neural Networks (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.