Project DetailsCryptographic systems, while mathematically secure, remain vulnerable to Side-Channel Analysis (SCA) attacks that exploit physical leakages such as power consumption and electromagnetic emissions. Existing deep learning–based SCA approaches are often impractical in real world attack scenarios, as they rely on implementation specific knowledge or require labeled internal masking data that is typically unavailable to a non invasive attacker. This project presents a masking agnostic, black box deep learning architecture capable of recovering secret keys from masked AES implementations without any prior knowledge of the masking scheme or access to internal intermediate values. Our approach utilizes a Multi-Task Learning (MTL) framework with a deep ResNet backbone to process raw traces and predict all key bytes simultaneously. To eliminate the dependency on hard coded, mask specific operations, we introduce a learnable bilinear combiner layer based on Canonical Polyadic (CP) decomposition. This layer enables the model to naturally learn the demasking operation directly from data, without explicit supervision. By integrating residual connections, L2 regularization, and Gaussian noise injection, the architecture ensures stable gradient flow and mitigates overfitting. Experimental results demonstrate that our masking-agnostic model achieves superior performance, successfully recovering all AES secret key bytes with perfect Guessing Entropy using as few as 2-4 traces on the ASCAD_r dataset with Boolean masking, significantly advancing the practicality of real-world SCA attacks. These results significantly advance the practical applicability of deep learning–based SCA in realistic black box attack settings.
To replicate the results, the following repository is provided: BlackBox DLSCA.
