Real-Time Multi-Modal Subcomponent-Level Measurements for Trustworthy System Monitoring and Malware Detection

Abstract

With increasingly sophisticated cyber-adversaries able to access a wider repertoire of mechanisms to implant malware such as ransomware, CPU/GPU keyloggers, and stealthy kernel rootkits, there is an urgent need for techniques to detect and mitigate such attacks. While state of the art relies on digital and analog side channel measurements assuming trustworthiness of measurements obtained on the main processor, such an approach has limitations since processor-based side channel measurements are potentially untrustworthy. Sophisticated adversaries (especially in late stage cyber attacks when they have breached the computer and network security systems such as firewalls and antivirus and penetrated the computer's OS) can compromise user-space and kernel-space measurements. To address this key limitation of state of the art, we propose a "subcomponent-level" approach to collect side channel measurements so as to enable robust anomaly detection in a modern computer even when the main processor is compromised. Our proposed approach leverages the fact that modern computers are complex systems with multiple interacting subcomponents and measurements from subcomponents can be used to detect anomalies even when the main processor is no longer trustworthy. We develop mechanisms to obtain time series measurements of activity of several subcomponents and methodologies to process and fuse these measurements for anomaly detection. The subcomponents include network interface controller, GPU, CPU Hardware Performance Counters, CPU power, and keyboard. Our main hypothesis is that subcomponent measurements can enable detection of security threats without requiring a trustworthy main processor. By enabling real-time measurements from multiple subcomponents, the goal is to provide a deeper visibility into system operation, thereby yielding a powerful tool to track system operation and detect anomalies.