Biologically-inspired Spiking Neural Networks (SNNs) are promising for energy-efficient neuromorphic computing in edge applications such as soft robotics, wearable health monitors, and IoT devices. Recent advances in flexible electronics enable deployment on thin, ultralow-cost conformal substrates. However, reduced packaging and thin encapsulation increase exposure to power side-channel leakage, while limited I/O and shared supply rails restrict conventional split-domain defenses. In this work, we present FlexSpy, a side-channel framework that integrates spike-accurate transient simulation with correlation and mutual information (MI) analysis, and introduces a Spike-Leakage Index (SLI) to localize vulnerable design blocks. Rather than targeting per-synapse weight extraction, the framework enables network-level recovery, label inference, and layer-wise spike-rate estimation. As a baseline, we compare against flexible recurrent neural networks (f-RNNs), which exhibit weaker leakage due to smoother recurrent dynamics. Across operating corners, FlexSpy achieves ROC-AUC = 0.91 and recovers layer-wise spike rates with NMSE = 0.14 and cosine similarity = 0.93. As countermeasure, spike-time randomization and event balancing are proposed to reduce leakage by 38–70% at ≤ 9% power overhead.
Sapui et al. (Thu,) studied this question.