Federated Learning and Self-Supervised Temporal Modeling for Collaborative Anomaly Detection in Distributed Tasks

Abstract

To address the challenges of highly concealed anomalies, complex cross-node relationships, significant temporal dependencies, and difficulties in centralized sharing of raw data in distributed task scenarios, this paper proposes a collaborative anomaly detection method based on federated learning and self-supervised temporal modeling. This method is designed for distributed task environments involving multiple nodes, multiple stages, and multiple sources of monitoring signals. Without uploading raw data, it achieves cross-client knowledge aggregation through a federated collaboration mechanism, alleviating the limitations imposed by data silos and privacy constraints on unified modeling. At the local modeling level, the method utilizes a self-supervised mask reconstruction strategy to mine potential temporal structure information from unlabeled execution sequences, enabling the model to learn contextual dependencies and dynamic patterns during task state evolution. Combined with temporal continuity constraints, it enhances the ability to identify complex behaviors such as anomalous perturbations, state shifts, and evolutionary instability. At the global optimization level, an aggregation mechanism oriented towards client-specific differences is constructed to improve the robustness and adaptability of the shared model in heterogeneous distributed environments. In the anomaly detection stage, an anomaly scoring method is established by integrating reconstruction deviations and temporal state change information, thereby improving the ability to perceive concealed and associated anomalies. This research focuses on anomaly data in open-source microservices. Results show that the proposed method effectively improves the accuracy and stability of anomaly detection in distributed tasks and demonstrates good applicability in complex operating environments. This study provides a feasible technical framework for distributed intelligent monitoring under privacy constraints and offers new methodological support for research on the operational status analysis and anomaly identification of complex systems.