Efficient, Portable Template Attacks

Abstract

Template attacks recover data values processed by tamper-resistant devices from side-channel waveforms, such as supply-current fluctuations (power analysis) or electromagnetic emissions. They first profile a device to generate multivariate statistics of the waveforms emitted for each of a set of known processed values, which then identify maximum-likelihood candidates of unknown processed values during an attack. We identify several practical obstacles arising in the implementation of template attacks, ranging from numerical errors to the incompatibility of templates across different devices, and propose and compare several solutions. We identify pooled covariance matrices and prior dimensionality reduction through Fisher's linear discriminant analysis as particularly efficient and effective, especially where many attack traces can be acquired. We evaluate alternative algorithms not only for the task of recovering key bytes from a hardware implementation of the Advanced Encryption Standard; we even reconstruct the value transferred by an individual byte-load instruction, with success rates reaching 85% (or a guessing entropy of less than a quarter bit remaining) after 1000 attack traces, thereby demonstrating direct eavesdropping of eight-bit parallel data lines. Using different devices during the profiling and attack phase can substantially reduce the effectiveness of template attacks. We demonstrate that the same problem can also occur across different measurement campaigns with the same device and that DC offsets (e.g., due to temperature drift) are a significant cause. We improve the portability of template parameters across devices by manipulating the DC content of the eigenvectors that form the projection matrix used for dimensionality reduction of the waveforms.