Multi-signal Anomaly Detection for Real-Time Embedded Systems

Abstract

This thesis presents MuSADET, an anomaly detection framework targeting timing anomalies found in event traces from real-time embedded systems. The method leverages stationary event generators, signal processing, and distance metrics to classify inter-arrival time sequences as normal/anomalous. Experimental evaluation of traces collected from two real-time embedded systems provides empirical evidence of MuSADET’s anomaly detection performance.

MuSADET is appropriate for embedded systems, where many event generators are intrinsically recurrent and generate stationary sequences of timestamp. To find timinganomalies, MuSADET compares the frequency domain features of an unknown trace to a normal model trained from well-behaved executions of the system. Each signal in the analysis trace receives a normal/anomalous score, which can help engineers isolate the source of the anomaly.

Empirical evidence of anomaly detection performed on traces collected from an industrygrade hexacopter and the Controller Area Network (CAN) bus deployed in a real vehicle demonstrates the feasibility of the proposed method. In all case studies, anomaly detection did not require an anomaly model while achieving high detection rates. For some of the studied scenarios, the true positive detection rate goes above 99 %, with false-positive rates below one %. The visualization of classification scores shows that some timing anomalies can propagate to multiple signals within the system. Comparison to the similar method, Signal Processing for Trace Analysis (SiPTA), indicates that MuSADET is superior in detection performance and provides complementary information that can help link anomalies to the process where they occurred.