A ThreshoId-ImpIementation-Based Neural-Network Accelerator Securing Model Parameters and Inputs Against Power Side-Channel Attacks

Neural network (NN) hardware accelerators are being widely deployed on low-power loT nodes for energy-efficient decision making. Embedded NN implementations can use locally stored proprietary models, and may operate over private inputs (e.g., health monitors with patient-specific biomedical classifiers [6]), which must not be disclosed. Side-channel attacks (SCA) are a major concern in embedded systems where physical access to the operating hardware can allow attackers to recover secret data by exploiting information leakage through power consumption, timing and electromagnetic emissions [1, 7, 8]. As shown in Fig. 34.3.1, SCA on embedded NN implementations can reveal the model parameters [9] as well as the inputs [10]. To address these concerns, we present an energy - efficient ASlC solution for protecting both the model parameters and the input data against power-based SCA.