Distributed Data Anomaly Detection Algorithm Integrating Federated Learning and Dynamic Representation Alignment Mechanism

Abstract

With the rapid development of distributed computing environments, IoT terminal networks, and cross-node data collaboration models, anomaly detection tasks face real challenges such as dispersed data storage, enhanced privacy constraints, and significant differences in features from multiple sources and heterogeneous structures. Traditional detection methods that rely on centralized aggregation of raw data can no longer meet the comprehensive requirements of security, collaboration, and robustness in complex scenarios. To address these issues, this paper proposes a distributed data anomaly detection method that integrates federated learning and dynamic representation alignment mechanisms, enabling multi-client collaborative modeling without directly sharing raw data. This method first uses a local encoding module to extract representations from the raw observation data of each node and combines this with anomaly scoring branches to enhance the model's ability to identify potential anomaly patterns. Subsequently, prototype regularization constraints are used to optimize the local feature space structure, improving the discriminative power between normal and anomaly samples. Based on this, a dynamic representation alignment mechanism is introduced to adaptively coordinate the semantic representations between different clients, mitigating feature shift and global aggregation distortion caused by data distribution heterogeneity. Finally, a federated aggregation strategy is used to achieve multi-node knowledge sharing and global anomaly detection model updates. This method integrates privacy protection, local discriminative modeling, and cross-node semantic consistency optimization into a single learning framework, enabling it to more effectively adapt to challenges such as sparse anomaly samples, significant node differences, and complex semantic drift in distributed scenarios. Experimental results demonstrate that the proposed method achieves superior performance across multiple evaluation metrics, validating the effectiveness and stability of the constructed framework in distributed anomaly detection tasks. Overall, this research provides a technically valuable solution for intelligent anomaly detection in privacy-constrained and heterogeneous collaborative environments.