ABSTRACT
This chapter investigates the vulnerabilities of SNNs against security threats. Novel attack methodologies and defensive countermeasures are proposed. The analyses are conducted both on discrete data, as well as on event-based data. Section 7.1 provides a comparative analysis of the security vulnerabilities of SNNs and DNNs with respect to the adversarial noise. Section 7.2 presents a cross-layer attack that threatens the SNNs’ robustness. A carefully crafted adversarial input noise triggers a hardware Trojan that injects bit-flips in the most vulnerable weight locations. Section 7.3 studies the inherent robustness of SNNs and explores different values of the SNNs’ structural parameters, which are the neuron's firing voltage threshold and time window boundary. Towards the SNNs’ security for event-based data, Section 7.4 presents a methodology for improving the robustness against adversarial attacks by employing noise filters for DVS sensors. Moreover, Section 7.5 presents a set of stealthy yet efficient adversarial attack methodologies targeted to perturb the event sequences and test them in the presence of noise filters for DVS cameras.
