Performance statistics and learning based detection of exploitative speculative attacks

Abstract

Most of the modern processors perform out-of-order speculative executions to maximise system performance. Spectre and Meltdown exploit these optimisations and execute certain instructions leading to leakage of confidential information of the victim. All the variants of this class of attacks necessarily exploit branch prediction or speculative execution. Using this insight, we develop a two step strategy to effectively detect these attacks using performance counter statistics, correlation coefficient model, deep neural network and fast Fourier transform. Our approach is expected to provide reliable, fast and highly accurate results with no perceivable loss in system performance or system overhead.