Machine Learning For Security: The Case of Side-Channel Attack Detection at Run-time

This paper presents experimental evaluation and comparative analysis on the use of various Machine Learning (ML) models for detecting Cache-based Side Channel Attacks (CSCAs) in Intel's x86 architecture. The paper provides performance evaluation of ML models based on run-time detection accuracy, speed, computational overhead, and distribution of error in terms of false positives and false negatives. Experiments are performed using state-of-the-art CSCAs namely; Flush+Reload and Flush+Flush attacks, under realistic load conditions on RSA and AES crypto-systems. The paper provides quantitative & qualitative analysis of at least 12 ML models being used for CSCA detection for the first time.