Spiking Neural Networks Vulnerable to Adversarial Attacks via Surrogate Gradients
A new study shared on arXiv reveals some serious flaws in spiking neural networks (SNNs) when they encounter adversarial examples. Researchers found that the success of white-box attacks on SNNs is heavily dependent on the choice of surrogate gradient estimators, even after the models have been trained against such attacks. By looking into how SNNs compare with Vision Transformers (ViTs) and CNNs, they identified two key problems: existing white-box attacks fail to use various surrogate gradient estimators for SNNs, and there's no single attack method that reliably tricks both SNNs and other models. This research sheds light on the vulnerabilities of SNNs, which are still catching up to conventional deep learning, emphasizing the need for better defenses in these efficient models.
Key facts
- arXiv:2209.03358v5 is a paper on SNN adversarial robustness.
- White-box attacks on SNNs depend on surrogate gradient estimators.
- Transferability of attacks across SNNs, ViTs, and CNNs is analyzed.
- No existing white-box attack uses multiple surrogate gradient estimators for SNNs.
- No single-model attack reliably fools both SNN and other architectures.
- SNN robustness study is underdeveloped compared to traditional deep learning.
- The paper advances the adversarial attack side of SNNs.
- Surrogate gradient estimators affect even adversarially trained SNNs.
Entities
Institutions
- arXiv