A Domain Generalization Methodology for Side-channel Attack

Side-channel attacks have been proven to be effective for retrieving secret keys by exploiting implementation vulnerabilities in encryption function. In recent years, deep learning techniques have been integrated into profiled side-channel attacks to achieve outstanding results. However, most state-of-the-art work only conducts attacks under ideal conditions where the profiling device (template device) and the target device are identical. The profiling traces are from the profiling device and the attack traces are from the target device. Unfortunately, constructing a template device is challenging and a flawed template device will lead to discrepancies between the profiling and the target devices. This inconsistent data distribution and discrepancies hamper the profiled SCA. To address these issues, we propose a domain generalization-based side-channel attack architecture called DGSCA. Especially, We simulate a scenario where the acquisition of target data is challenging and can’t be used for model training. Initially, we combine the Deep CORAL method with Mixup to deserve domain invariant representations across source domains. Subsequently, we integrate meta-learning into the DGSCA framework to enhance the model’s generalization ability, which is accomplished by randomly partitioning the source and target domains and leveraging the knowledge gained from the source domain to improve the attack result in the target domain. Moreover, we implement four different deep learning models (MLP, CNN, ResNet, DenseNet) and evaluate the DGSCA framework using the publicly available ASCAD dataset. The experimental results on the ASCAD desynchronous dataset show that our method successfully recovers the key within 1665 traces and demonstrate the scalability and generality of our proposed framework.