Fooling Deep Neural Networks: A Strategyfor Mitigating DLSCA

Side-channel analysis (SCA), particularly deep learning-based SCA (DLSCA), threatens cryptographic implementations by automatically exploiting complex and nonlinear leakage and weakening countermeasures such as masking. Existing countermeasures against DLSCA mainly suppress, decorrelate, or obscure exploitable leakage. This work introduces Spoofing-based Mitigation (SPM), a leakage-shaping approach designed to mislead key-recovery attacks by biasing their key hypotheses. SPM introduces a controllable offset into AES intermediate-value computations and compensates for it to preserve the correct ciphertext, while shaping the relationship between measured leakage and the intermediate-value labels used for profiling. We implement SPM at the first-round S-box output and evaluate it at first-round and last-round attack points on an XMEGA128D4 microcontroller using a ChipWhisperer platform. Under the evaluated identity-label settings, the induced key’s rank rapidly reaches zero across seven CNN-based DLSCA models, two MLP-based models, and a template attack, while the correct key remains unrecovered. These results provide experimental evidence that leakage shaping can misdirect the evaluated key-recovery attacks. The collected dataset is publicly available to support reproducible research.