Secure-Aggregation Federated Graph Attention Networks with SHAP-Based Edge Attribution for Advanced Persistent Threat Detection

Advanced Persistent Threats (APTs) pose significant cybersecurity challenges due to their stealthy, prolonged activities and the limited visibility of malicious behaviour in host provenance data. Existing detection approaches often require centralized access to sensitive logs and provide limited interpretability, raising concerns about data confidentiality and analyst trust. To address these challenges, this study proposes a privacy-preserving and explainable federated learning framework for APT detection using provenance graphs. The proposed framework integrates edge-typed Graph Attention Networks (GATv2), self-supervised masked-feature learning, and federated prototype-based anomaly detection to identify suspicious activities without requiring labeled attack data during training. To preserve confidentiality, a dropout-resilient, Bonawitz-style secure aggregation protocol is employed to protect model updates and shared statistics, including normalization parameters, clustering information, and calibration distributions. Furthermore, edge-level Shapley value attribution is incorporated to provide interpretable explanations of detected anomalies and support security analysts during incident investigation. The framework is evaluated using the DARPA Transparent Computing Engagement 3 benchmark, specifically the THEIA and CADETS datasets, under a simulated four-client federated setting. Experimental results across three random seeds demonstrate strong detection performance, achieving Area Under the Receiver Operating Characteristic Curve (AUROC) values of 0.995 and 0.971 and Area Under the Precision–Recall Curve (AUPRC) values of 0.870 and 0.561 for THEIA and CADETS, respectively. The proposed secure aggregation mechanism maintains detection performance comparable to conventional federated averaging, introduces approximately 0.1% additional training time on THEIA, and reduces cross-round client-linkage attack accuracy from 85.6% to chance-level performance. Behaviour-level analysis demonstrates complete detection of scan-like attack activities at a calibrated 95% threshold, while other malicious activities exhibit threshold-dependent recall. The explainability analysis confirms the completeness of Shapley attributions and convergence of the highest-ranked contributing edges within 25 permutations, although their empirical fidelity is comparable to less computationally expensive explanation methods. Overall, the findings demonstrate the potential of combining secure federated learning, provenance-graph representation learning, and explainable AI for confidentiality-preserving APT detection, while highlighting important limitations related to threshold sensitivity, explanation fidelity, and semi-honest security assumptions.