SNR to Success Rate: Reaching the Limit of Non-Profiling DPA

Many profiling power analysis attacks estimate the multivariate probability distribution using a profiling step, and thus, can optimally combine the leakages of multiple sample points. Though there exists several approaches like filtering, Principal Component Analysis for combining the leakages of multiple sample points in non-profiling DPA, their optimality has been been rarely studied. We study the issue of optimally combining the leakages of multiple sample points in non-profiling DPA attacks using a linear function. In this work, our contributions are three-fold: 1) we first derive a relation between the success rate of a CPA attack and the SNR of the power traces, 2) we introduce a multivariate leakage model for Virtex-5 FPGA device, and 3) using the proposed multivariate leakage model, we devise linear filters to maximize the SNR of the output leakage which, in turn, optimizes the success rate of the CPA attacks in a non-profiling setup.