Robust and Efficient Adversarial Defense in SNNs via Image Purification and Joint Detection

Abstract

Spiking neural networks (SNNs) leverage neural spikes to provide solutions for low-power intelligent applications on neuromorphic hardware. Although the spiking mechanism significantly enhances computational efficiency, especially in energy-constrained environments, they still lack resistance to noise perturbations and adversarial attacks. In this paper, we propose a defense framework based entirely on SNNs and design a fast and efficient training algorithm. The framework is divided into an image purification module and an adversarial detection module. The image purification module is employed for the extraction of noise and the reconstruction of input images. The adversarial detection module is utilized to differentiate between clean and adversarial images, thereby further enhancing defense performance. Meanwhile, our approach is highly flexible and can be seamlessly integrated with other defense strategies. Experimental results demonstrate that the proposed methodology outperforms state-of-the-art baselines in terms of defense effectiveness, training time and resource consumption.