A Desynchronization-Based Countermeasure Against Side-Channel Analysis of Neural Networks

Model extraction attacks have been widely applied, which can normally be used to recover confidential parameters of neural networks for multiple layers. Recently, side-channel analysis allows parameter even with several deep layers high effectiveness. It is therefore interest implement a certain level protection against these attacks. In this paper, we propose desynchronization-based countermeasure that makes the timing activation functions harder. We analyze properties and design desynchronization in way dependency on input type hidden. experimentally verify effectiveness 32-bit ARM Cortex-M4 microcontroller employ t-test show information leakage. The overhead ultimately depends number neurons fully-connected layer, example, case 4096 VGG-19, overheads are between 2.8% 11%.