Stacked ensemble models evaluation on DL based SCA

Side-channel analysis (SCA) has proved its effectiveness in attacking the cryptographic implementations for high security algorithms like Advanced Encryption Standard (AES). The improvement of machine learning (ML) in general and deep learning (DL) in particular in SCA shows that DL is a big threat in hardware security, in which the secret key of SCA countermeasure AES can be revealed with only 40 traces [17]. Combination of multiple DL can improve the effectiveness of SCA more. However, how to combine those models together is still a question. This paper applied stacked ensemble ML to combine the predictions from a number of inputs and sub-models together to enhance the power of DL in attacking AES implementation with SCA countermeasures. We train not only the output probabilities of sub-models but also their maximum likelihood score (MLS) in a stacked ensemble model to improve the performance of convolutional neural network (CNN)-based models. Further more, output probabilities from multiple trace inputs to the same sub-model are also utilized for our stack ensemble model training. A two step training procedure is required, one is for each sub-model and the other is for the stacked ensemble model, which takes inputs as the output probabilities of the above trained sub-models. This paper evaluate the effectiveness of various stacked ensemble models in terms of the number of input traces and the number of sub-models used for the first training stage. Our best model generates state-of-the art results when attacking the ASCAD variable-key database, which has a restricted number of training traces per key, recovering the key within 20 attack traces [15] in comparison to 40 traces as required by the original CNN with Plaintext feature extension (CNNP)-based model.