RESEARCH · RESEARCH · #2704
Benchmark study finds spiking neural networks harder to steal than ANNs
arXiv:2610.36298v1 benchmarks four representative model-stealing attacks (Papernot, Tramèr, T-RND, color perturbation) against spiking neural networks (SNNs) and comparable artificial neural networks (ANNs) on MNIST, Fashion-MNIST, and CIFAR-10. Across all attacks and datasets, SNNs were consistently harder to replicate: surrogate models achieved roughly 10–20% lower agreement with SNN targets and 15–25% lower adversarial transferability than surrogates of ANNs, a difference the authors attribute to SNNs' spike-based, time-dependent computation and resulting decision boundaries.
KEY POINTS
- arXiv:2610.36298v1 benchmarks four representative model-stealing attacks (Papernot, Tramèr, T-RND, color perturbation) against spiking neural networks (SNNs) and comparable artificial neural networks (ANNs) on MNIST, Fashion-MNIST, and CIFAR-10.
- Across all attacks and datasets, SNNs were consistently harder to replicate: surrogate models achieved roughly 10–20% lower agreement with SNN targets and 15–25% lower adversarial transferability than surrogates of ANNs, a difference the authors attribute to SNNs' spike-based, time-dependent computation and resulting decision boundaries.
- If reproducibility of model behavior from black-box queries is genuinely harder for SNNs, neuromorphic architectures could offer a practical security advantage for deployed models in sensitive settings.
WHY IT MATTERS
If reproducibility of model behavior from black-box queries is genuinely harder for SNNs, neuromorphic architectures could offer a practical security advantage for deployed models in sensitive settings.