Abstract
Mapping specific sections of Android malware code to high-level behavioral tactics from the MITRE ATT&CK framework is traditionally a manual, labor-intensive task. This paper presents an automated pipeline that combines a novel graph representation technique, SIR-GN (Structural Iterative Representation for Graph Nodes), with explainable AI (SHAP) to identify and localize malicious tactics within Control Flow Graphs (CFGs). Our results demonstrate that SIR-GN significantly outperforms standard Graph Neural Networks, providing both high accuracy and actionable interpretability for security analysts.
Introduction
Security researchers rely on the MITRE ATT&CK ontology to categorize adversary behaviors. However, identifying which specific code segments in an Android application correspond to tactics like "Initial Access" or "Defense Evasion" is challenging due to the scale and complexity of Control Flow Graphs (CFGs). Current automated tools often focus on binary classification (malicious vs. benign) rather than granular tactic attribution. This research addresses this gap by proposing a method to pinpoint malicious subgraphs within raw executables.
Methodology
The proposed system utilizes a three-stage pipeline:
- Graph Extraction: Static analysis is performed on Android APKs using AndroGuard to generate directed CFGs.
- Representation Learning (SIR-GN): Instead of standard connectivity-based Graph Neural Networks (GNNs), the authors employ SIR-GN. This method generates node representations based on structural similarity, creating a permutation-invariant "pseudo adjacency matrix."
- Classification and Attribution: A Random Forest classifier identifies Tactics, Techniques, and Procedures (TTPs). To provide interpretability, SHAP values are propagated back through the graph representation to highlight the specific edges and nodes responsible for the classification.
Experimental Results
The model was evaluated on a dataset of 3,250 Android malware samples from VirusTotal, encompassing over 136,000 nodes. The proposed SIR-GN method was compared against standard GNN architectures, including Graph Isomorphic Networks (GIN) and Graph Attention Networks (GAT).
As shown in the table above, SIR-GN achieved an F1 score of 0.927 and an accuracy of 0.896, representing a substantial improvement over GIN and GAT. Qualitative analysis further validated the model; for instance, it correctly identified API calls related to malicious ad-networks (e.g., com/madhouse/android/ads) as indicators of the "Initial Access" tactic.
Discussion
The success of SIR-GN highlights that the structural "shape" of program logic—such as specific branching and looping patterns—is a more potent indicator of malicious intent than simple node connectivity. Furthermore, the integration of SHAP allows the model to act as a "searchlight," localizing specific malicious "needles" within the vast "haystack" of an application's control flow.
Conclusion and Future Work
While the SIR-GN approach shows great promise, challenges remain regarding data imbalance in certain TTP categories and the potential for code obfuscation to hinder static analysis. Future research will focus on scaling the dataset and exploring dynamic analysis techniques to complement the current static graph representation.
Summary updated