Mind and Body: A Mini Perspective in Software-Hardware Interconnection in Neuromorphic Computing Security

Neuromorphic computing offers a highly efficient paradigm that relies on discrete events, contrasting heavily with traditional sequential processing. However, it remains critically underexplored at the crossover of hardware and software security. To address this gap, this paper conducts a mini analysis in the specific attack vectors and defensive strategies inherent to neuromorphic computing. By examining Spiking Neural Networks and Spiking Deep Belief Networks, the research highlights significant vulnerabilities in the software layer. These include the high transferability of adversarial examples from traditional neural networks and a distinct susceptibility to engineered perturbations. Furthermore, physical hardware vulnerabilities are analyzed, demonstrating how properties such as voltage manipulation and bit manipulation in FPGA circuits introduce severe risks. These risks include hardware trojans triggered by inputs and side channel attacks. In response, physical and logical layer defenses are evaluated for their efficacy in active hardware guarding. These defenses include robust driver designs acting as voltage comparators and the integration of dummy neurons for continuous monitoring. Ultimately, this perspective emphasizes a broad thematic interconnection between software algorithms and hardware constraints. Securing neuromorphic systems requires creative defenses to protect physical pathways without compromising the inherent functionality of computing inspired by the brain.