A deep learning based side-channel attack on BIKE

Abstract Post-quantum cryptographic implementations are increasingly exposed to side-channel threats, making implementation-level security assessment an essential complement to algorithmic analysis. BIKE is a representative code-based key encapsulation mechanism, and its latest specification adopts the Black-Gray-Flip (BGF) decoder, which offers practical advantages in decoding performance and implementation regularity. However, the side-channel security of this decoding scheme has not been adequately studied in the latest BIKE reference implementation. In this work, we investigate this problem and develop a deep-learning-based side-channel attack for recovering a private-key-dependent sequence from long BIKE decoding traces. Firstly, we examine the BGF decoding procedure in the latest BIKE reference implementation and show that its repeated counter computations and threshold-based bit-flipping decisions give rise to exploitable side-channel leakage. Secondly, we transform BIKE BGF decoding leakage recovery into an alignment-free sequence-to-sequence task and instantiate it with a convolutional recurrent neural network (CRNN) trained using Connectionist Temporal Classification (CTC). This formulation enables direct recovery of the target sequence from complete traces without relying on manually defined temporal boundaries. Finally, we evaluate the proposed formulation through end-to-end recovery experiments, ablation analysis, and comparison with representative baselines. In the evaluated profiled, unprotected, same-platform ARM Cortex-M4 setting, experimental results show that the proposed method achieves 99.69% bit accuracy and 95.68% perfect sequence recovery on the independent test set. Our target is a 16-bit sequence of counter-comparison outcomes generated during BIKE decoding; this sequence is a private-key-dependent intermediate rather than the complete private key. Complete private-key recovery would require accumulating such observations over multiple executions and applying an additional post-processing stage. Overall, these results show that alignment-free sequence recovery is effective for BIKE side-channel analysis on long, temporally unstable traces in this evaluated setting.