Brain-inspired Lab Peking University

Neuromorphic Computing Paradigms Enhance Robustness through Spiking Neural Networks

spiking neural networks adversarial robustness neuromorphic computing

Jianhao Ding, Zhaofei Yu, Jian K. Liu, Tiejun Huang. Neuromorphic Computing Paradigms Enhance Robustness through Spiking Neural Networks. Nature Communications. 2025, 16(1): 10175.*

Read the paper

Research Background and Problem

Background

Deep learning models have achieved remarkable success across a wide range of tasks, including image recognition and decision-making systems, but their reliability is still threatened by adversarial attacks that introduce imperceptible perturbations to input data.

Limitation of Existing Methods

Existing artificial neural networks (ANNs) exhibit limited robustness, as even minor input modifications can lead to significantly incorrect predictions, which raises serious concerns in safety-critical applications such as autonomous driving and human-machine interaction.

Our Perspective

To address this limitation, we explore neuromorphic computing paradigms and leverage spiking neural networks (SNNs), which exploit temporal dynamics and brain-inspired processing mechanisms to enhance robustness against adversarial perturbations.

figure image

Contributions

  • Demonstrate that temporal processing in SNNs is a key factor for improving robustness against adversarial attacks.
  • Propose encoding strategies that prioritize task-critical information early in the temporal sequence.
  • Introduce early-exit decoding methods to mitigate the impact of later perturbations.
  • Design specialized training algorithms to better capture temporal dependencies and improve generalization.
  • Propose a fusion encoding strategy to balance robustness and performance on natural data.

Core Method

The proposed approach leverages the temporal characteristics of spiking neural networks to improve robustness. Specifically, input data are encoded into temporal spike sequences where important information is presented earlier, allowing the network to make reliable decisions before adversarial perturbations accumulate. In addition, early-exit decoding mechanisms reduce the influence of later noise, while temporal-aware training algorithms optimize spike timing dependencies. Furthermore, a fusion encoding strategy integrates multiple encoding schemes to enhance robustness under diverse attack scenarios.

SNNs exhibit capacities to protect themselves against adversarial attacks imperceptible to human vision by leveraging meticulously crafted encoding strategies and precise control over decoding duration.

figure image

Representative Results

Experimental results on datasets such as CIFAR-10 show that the proposed SNN-based methods significantly outperform conventional ANNs in adversarial settings, achieving up to twice the robustness. At the same time, the model maintains high energy efficiency, making it suitable for real-world safety-critical and resource-constrained applications.

Accuracy comparison of SNNs with the fused encoding (enc.) versus non-fused ANN and SNN models, using the rate and TTFS decodings (dec.) on clean and four attacked datasets.
Accuracy comparison of SNNs with the fused encoding (enc.) versus non-fused ANN and SNN models, using the rate and TTFS decodings (dec.) on clean and four attacked datasets.