Efficient CNN-Based Side-Channel Attacks on Dilithium without Device Access

The post-quantum cryptography standard, released in August 2024, faces significant threats from side-channel attacks (SCAs). While non-profiled attacks demand extensive traces and time, profiled attacks, though more effective, require access to identical devices. This paper proposes a hybrid SCA on Dilithium, leveraging the strengths of both approaches. Our method profiles a template using leakage from operations with known variables, then applies it to target sensitive variables. Specifically, Convolutional Neural Networks (CNNs) are employed to model leakages from Inverse Number Theoretic Transform (INTT) operations. This model is subsequently used to recover private keys by targeting INTT operations involving c. Implemented on an ARM Cortex-M4 platform, our attack requires only 10/27/18 power traces to recover keys for Dilithium2/3/5, respectively, significantly reducing the number of traces compared to prior non-profiled attacks requiring hundreds.