Invited Paper: Toward Secure In-Sensor Intelligence: Threats and Defenses in SNNs

Spiking Neural Networks (SNNs) are inspired by the event-driven and temporally sparse nature of biological neurons, enabling deployment in in-sensor computing systems. The sensing and computation being tightly coupled in the in-sensor devices help in low-latency and energy-efficient data processing. This paradigm introduces a novel security front, exposing vulnerabilities in both neuromorphic hardware and the temporally sparse spike encodings to a variety of emerging attack modalities. This survey offers a comprehensive examination of the security and robustness landscape for SNNs deployed in in-sensor computing environments. It begins by outlining the architectural and algorithmic characteristics that define in-sensor SNN pipelines, with particular focus on temporal coding, asynchronous processing, and hardware constraints. We then review pertinent threat models, including spike-level adversarial perturbations, sensor spoofing, electromagnetic interference, fault injection, and timing-based privacy leakage, considering both white-box and black-box attack scenarios that exploit spatiotemporal vulnerabilities. Existing defense mechanisms, spanning noise shaping, homeostatic control, adversarial training, secure spike encoding, and hardware-level protections, are systematically categorized and assessed in the context of resource-constrained, event-driven platforms. Finally, we highlight emerging research directions in secure neuromorphic learning, such as continual and federated SNN training under adversarial settings, establishing a foundation for advancing the research in secure neuromorphic systems.