Research On Optimization Methods Of Side Channel Profiled Attacks Based On Deep Learning

As a powerful attack way,side channel attacks have become a significant part in the field of cryptanalysis.They exploit the cryptographic algorithm side-channel leaks at runtime of encrypted devices for cryptanalysis,among which profiled side channel attacks have attracted the attention of many scholars due to the extraordinary performance.Recently,deep learning has been introduced to the profiled side channel attacks and outperforms traditional methods.Despite the great success,there still exist many aspects that need to be further optimized for learning-based side channel attacks.For the ideal attack scenario,most works only infer the final results from just a single view but neglect to exploit the information from multiple views simultaneously.For the actual attack scenario,the attack performance will be degraded due to the domain shift between the profiling devices and the attack devices.Aiming at the shortcomings of learning-based profiled side channel attacks under the ideal attack scenario and the actual attack scenario,this thesis proposes optimization methods.Details are as follows:1.For the ideal attack scenario,i.e.when the profiling set and the attack set are from the same device,this thesis proposes a multi-view information aggregation optimization method named MV-Net for profiled side channel attacks.We first image the side channel trace using Gramian Summation Angular Field and Gramian Difference Angular Field respectively.Then we obtain the two-channel converted image by concatenating the Gramian Summation Angular Field image and the Gramian Difference Angular Field image along the channel dimension.Regarding the Gramian Angular Field image and the original trace as two views,we utilize compact multimodal bilinear pooling to fuse the heterogeneous features extracted from the two views.Experimental results on benchmark datasets demonstrate the superior performance of ours compared with prior methods using single-view information.Besides,we analyze the good performance of MV-Net by visualizing the attention maps.The visualization results illustrate that MV-Net can effectively focus on the informative regions of the trace and the converted Gramian Angular Field image.2.For the actual attack scenario,i.e.when the profiling set and the attack set are from different devices,this thesis proposes an optimization method aimed at crossdevice profiled side channel attacks based on mixup regularized adversarial domain adaptation.During the profiling phase,we utilize the training data collected from the cloned device to train a network.Then in the adaptation phase,we utilize a small number of unlabeled traces collected from the target device to fine-tune the pre-trained network.In the attack phase,we utilize the fine-tuned network to attack the target device.Moreover,in order to improve the generalization of the network,we apply the mixup regularization during the adaptation process.Mixup can achieve the goal of data augmentation by constructing the nearest neighbor samples of training samples,and thus can improve the discriminability of the class-aware information while learning the domain-invariant features across domains.The experimental results on benchmark datasets illustrate that our proposed method can largely improve the attack performance under the actual attack scenario.