Raw event-based adversarial attacks for Spiking Neural Networks with configurable latencies.

Abstract

Spiking Neural Networks (SNNs) and data from Dynamic Vision Sensors (DVSs) offer energy-efficient solutions for edge devices with limited battery life. The input latencies of event data to SNNs are critical for energy savings, and reducing these latencies through configurable parameters is essential. However, security concerns, particularly adversarial attacks on SNNs, are increasingly significant. While most existing research primarily focuses on attacking event sequences, which may not always be accessible, attacks on raw event streams with configurable latencies remain underexplored due to challenges such as extreme sparsity and the complexity of discrete optimization. This paper proposes a novel adversarial attack method on raw event streams with configurable latencies. To address sparsity and discrete optimization, we smooth the optimization by converting binary spikes into continuous values. Furthermore, we introduce an adaptively stochastic strategy for sampling attacking latencies. We apply regularization terms to maintain sparsity and ensure adversarial samples resemble raw event streams while approximating the target label. Extensive experiments on datasets such as N-MNIST, CIFAR10-DVS, N-Caltech-101, and Gesture-DVS demonstrate that our method consistently outperforms existing approaches, achieving higher attack success rates (ASR) across various latencies. Ablation studies validate the effectiveness of our contributions and highlight the impact of latency on the generation of adversarial samples.