Switch Capacitor-Based Time-Varying Transfer Function for FCN and CNN MLSCA-Resistant AES256 in 65-nm CMOS

Mathematically secure cryptographic implementations can leak critical information through physical side channels. Machine learning (ML) has facilitated efficient side-channel analysis (SCA), especially on small IoT devices and smart cards. We propose a lightweight, synthesizable technique to enhance ML-based SCA resilience. Our approach introduces a physical time variance technique that specifically targets Deep Neural Network based MLSCA. This brief presents a physical time variance technique that is effective against CNN contrary to the previous state-of-the-art. By eliminating analog units and utilizing a switched capacitor design, it outperforms existing techniques by 5x in terms of traces to train the attacking neural network.