Context-Aware Detection and Resolution of Data Anomalies for Semi-Autonomous Cyber-Physical Systems

Abstract

A cyber-physical system (CPS) with both autonomous and manual control capabilities, or a semi-autonomous (SA) system, is one of the most commonly seen types of system in our daily lives, such as cars, airplanes and ships. While having the benefits of autonomous control to enhance safety/comfort of transportation and the flexibility of manual control to handle safety-critical situations, SA systems inevitably inherit the vulnerabilities embedded in both control types. That is, an SA system will also suffer from component failures or design/software bugs (e.g., crashes of Boeing 737 MAX) and potential attacks (e.g., sensor spoofing) as a general CPS does. Moreover, since mechanical components are gradually being replaced by their electronic counterparts in SA systems, this trend also introduces new reliability and security risks — increasing adoption of multiple heterogeneous communication interfaces widens attack surfaces that an adversary can exploit.

Considering the potential security and safety concerns caused by system faults/flaws, human error, and malicious attacks, we develop a suite of mechanisms/systems for detection and resolution of system anomalies by cross-validating the sensor data and the context information to enhance the security and safety of SA systems from three key perspectives that can directly influence the operation of SA systems — system operation, received information, and control decisions. In this thesis, we propose both domain-general design for SA systems and its domain-specific realization using SA vehicles as a concrete case study. From the system operation perspective, we propose CADD, a context-aware anomaly detection system, to capture abnormal system behavior under various operation contexts while considering practical scenarios where some (context) information cannot be observed by the SA system. We then present DiVa, a diagnostic system that pinpoints/identifies the anomalous source(s) after an anomaly is detected. It exploits the cyber-physical correlation or causality between the internal data of the SA system to narrow down the origin of anomaly while assuming no data can be entirely trusted. From the received information perspective, we propose EDRoad, an easy-to-use system for verification of the data received from external sources to ensure no compromised data will be used to provide services to the SA systems. Finally, from the control decision perspective, we introduce CADCA, a system for detecting and resolving control actions that may potentially lead to unstable system states or safety-critical situations.