Differential Metric Based Deep Learning Methodology for Non-Profiled Side Channel Analysis

Abstract

Power Side-Channel analysis recovers sensitive information not only from physical proximity to a device but also from basic knowledge of sample leaked data collection. With minimum mean squared error metric, power analysis using a deep learning test case increases the confidence level of proper identification of leaked data. Comparison with state-of-the-art technology in this work shows improved performance in the non-profiled SCA category of detection. The deep learning technique aids in calculating the average loss gradient values and the loss values, both being calculated by taking the traces in MathWorks implementation as the training data and the MSB values of the intermediate values as the training labels to reveal the expected secret key. Moreover, iterative training of some machine learning techniques with different FPGA boards implementing cryptographic designs increased the accuracy of leakage detection at an earlier stage to a better extent.