Towards reducing costs of side channel analysis for real time algorithm detection

Side channel analysis using asynchronous sampling can be used to detect algorithms in real-time, but the cost of performing it in an automated manner is currently too expensive to implement in a small and low-cost form factor. In this work, we show that it is possible to reduce the performance overhead of detecting specific algorithms running on a target processor by reducing the need to identify individual instructions with high accuracy. We show that by using features that correspond loosely to instructions, combined with a lightweight sliding window approach, algorithms can be distinguished even though individual instructions may not be identified with high certainty. We explore the tradeoffs in window size, feature selection, and sampling rate. Our experimental results demonstrate comparable detection performance to a standard unconstrained approach, demonstrating that lightweight side-channel analysis is feasible for automated, real-time security applications.