Adversarial Attacks and Defense Mechanisms in Spiking Neural Networks: A Comprehensive Review
Bahareh Kaviani Baghbaderani, Hossein Shahinzadeh
Abstract
Spiking Neural Networks (SNNs), often referred to as third-generation neural networks, are inspired by the workings of the human brain and are celebrated for their energy efficiency, positioning them as a promising alternative to traditional Artificial Neural Networks (ANNs). However, SNNs face significant challenges in robustness, particularly under adversarial attacks, which can undermine their performance in real-world applications. This survey provides a detailed examination of adversarial attacks targeting SNNs, categorizing existing attack methods and defense mechanisms while highlighting inherent structural characteristics and training approaches that influence robustness. Unlike prior surveys that focus primarily on adversarial attacks in traditional ANNs, this work is the first to consolidate and critically analyze adversarial strategies and defenses specific to SNNs. Furthermore, we quantitatively compare SNNs with ANNs in terms of their susceptibility to adversarial attacks, offering new insights into the unique challenges and opportunities presented by SNNs. These contributions establish this survey as a foundational reference for advancing robust SNN research and development.