Deep Learning Side-Channel Attack Resilient AES-256 using Current Domain Signature Attenuation in 65nm CMOS
Debayan Das, Josef Danial, Anupam Golder, Santosh Ghosh, Arijit Raycho Wdhury, Shreyas Sen · 2020
This article, for the first time, demonstrates an efficient circuit-level countermeasure to prevent deep-learning based side-channel analysis (DLSCA) attacks on encryption devices. Machine learning (ML) SCA, particularly DLSCA attacks have been shown to be extremely effective as it can potentially reveal the secret key of the cryptographic device with as low as a single trace, by offloading the heavy-lifting on the profiling phase where the model learns the correlated leakage patterns of the key. This work presents a current-domain signature attenuation (CDSA) hardware embedding an AES256 engine fabricated in 65nm CMOS technology to suppress the current signature by >350× before it reaches the power supply pin accessible to an attacker. Measurement results show that a 256-class deep neural network (DNN) model for DLSCA attack can be fully trained (>99.9% test accuracy) using only <; 5K power traces from the unprotected AES256, while the DNN model for the protected CDSA-AES256 cannot be trained even with 10M traces.