Analysis Of Side-channel Attack On Embedded Devices Based On Machine Learning

Electromagnetic leakage signal is the electromagnetic radiation signal produced by the change of internal current during the normal operation of electronic system.Under normal circumstances,electromagnetic leakage signals are observed at multiple locations such as resistors,power supplies,and computing chips.This article classifies and recognizes the electromagnetic leakage signals generated by embedded devices,which is of great significance for evaluating the safety of embedded devices.This article has four main innovations as follows:Firstly,we built an automatic acquisition platform for electromagnetic leakage signals from the storage signals of the single-chip microcomputer and SIMECK32/64 signals,and an automatic acquisition platform for electromagnetic leakage signals from the neural network of Jetson Nano.Corresponding hardware components and software are designed to make the collection process and data screening process more efficient.Secondly,in order to solve the problem that the gradient of the direct training CNN does not decrease,a new CNN structure,Easy Task Network,is proposed based on the theory of Hamming Weight Model.And on the basis of this network structure,an accuracy of 96.80%is achieved on the storage task data set.Thirdly,without knowing any plaintext,ciphertext,and leaked intermediate value information,use a fully random key to encrypt with the SIMECK32/64 algorithm,carry out electromagnetic leakage attacks,and use a customized neural network structure with tag Tansfer Learning,reached an average accuracy of 96%.In addition,transfer learning is carried out through data sets of different one-dimensional electromagnetic leakage signals,which improves the accuracy rate by up to 6%,which illustrates the transferability between one-dimensional electromagnetic leakage signals of different tasks.Fourthly,conduct side-channel attacks based on electromagnetic leakage against Jetson equipment.Two-dimensional security assessments were conducted,one was to classify the leakage signals of seven different neural networks under the same framework TensorRT,and the other was to classify the electromagnetic leakage of the same neural network under the three frameworks of TensorRT,Tensorflow,and Pytorch.In this task,we designed and tested the structure of how to convert a two-dimensional convolutional neural network into a one-dimensional convolutional neural network through different strategies.The test verified its feasibility and proved that a good network conversion strategy can improve the network by 5%?12%Classification recognition accuracy.