Federated Multi-Scale Representation Learning for Privacy-Aware Log Anomaly Detection in Distributed Cloud Environments

Abstract

This work addresses scenarios where multiple organizations operate cloud native systems and face sensitive log data, restricted sharing, and cross-domain anomaly propagation that is difficult to characterize. It proposes a federated representation-based framework for log anomaly identification. Each participant structures and semantically encodes its local log sequence to produce low-dimensional latent features, and dynamic dependencies of system behavior are captured through multi-scale temporal aggregation with sliding windows. In cross-domain collaboration, federated updates and parameter aggregation enable knowledge sharing without exposing raw logs, while a representation-level consistency constraint mitigates distribution differences and semantic drift across clients, improving stability and transferability of the shared subspace. For anomaly modeling, the method constructs a distance-driven local risk scoring strategy using the normal semantic center as reference, providing interpretable abnormal discrimination even under weak or missing labels. Heterogeneous multi-client settings are built on a unified public log benchmark, and systematic sensitivity evaluation on optimizer choice, representation dimension, anomaly proportion, and template noise demonstrates stable performance and strong cross-domain robustness under diverse perturbations.