Advanced Threat Intelligence: Interpretation of Anomalous Behavior in Ubiquitous Kernel Processes
Targeted attacks on digital infrastructures are a rising threat against the confidentiality, integrity, and availability of both IT systems and sensitive data. With the emergence of advanced persistent threats (APTs), identifying and understanding such attacks has become an increasingly difficult task. Current signature-based systems are heavily reliant on fixed patterns that struggle with unknown or evasive applications, while behavior-based solutions usually leave most of the interpretative work to a human analyst. This thesis presents a multi-stage system able to detect and classify anomalous behavior within a user session by observing and analyzing ubiquitous kernel processes. Application candidates suitable for monitoring are initially selected through an adapted sentiment mining process using a score based on the log likelihood ratio (LLR). For transparent anomaly detection within a corpus of associated events, the author utilizes star structures, a bipartite representation designed to approximate the edit distance between graphs. Templates describing nominal behavior are generated automatically and are used for the computation of both an anomaly score and a report containing all deviating events. The extracted anomalies are classified using the Random Forest (RF) and Support Vector Machine (SVM) algorithms. Ultimately, the newly labeled patterns are mapped to a dedicated APT attacker–defender model that considers objectives, actions, actors, as well as assets, thereby bridging the gap between attack indicators and detailed threat semantics. This enables both risk assessment and decision support for mitigating targeted attacks. Results show that the prototype system is capable of identifying 99.8% of all star structure anomalies as benign or malicious. In multi-class scenarios that seek to associate each anomaly with a distinct attack pattern belonging to a particular APT stage we achieve a solid accuracy of 95.7%. Furthermore, we demonstrate that 88.3% of observed attacks could be identified by analyzing and classifying a single ubiquitous Windows process for a mere 10 seconds, thereby eliminating the necessity to monitor each and every (unknown) application running on a system. With its semantic take on threat detection and classification, the proposed system offers a formal as well as technical solution to an information security challenge of great significance.
- PhD