SplitAD: A lightweight and privacy-enhancing vertical federated anomaly detection framework based on hierarchical autoencoders

With the explosive growth of data volume, coupled with the proposal of the federated learning paradigm, federated anomaly detection , with the aim of extracting features and identifying anomalies from distributed data sources , has attracted extensive concerns. However, privacy leakage and vertical data distribution remain major challenges in federated anomaly detection, and existing schemes struggle to balance privacy and cost. Therefore, this paper introduces SplitAD, a lightweight and privacy-enhancing vertical federated anomaly detection framework that innovatively leverages hierarchical autoencoders . Specifically, SplitAD assigns the bottom and top autoencoders to different participants, where the bottom one is held by each participant to learn and extract patterns from its local data, and the top one is used to detect anomalies according to all participants' reconstruction errors. Then, a series of model training and execution algorithms are designed for collaborative anomaly detection. Furthermore, since only perturbed reconstruction errors are shared, compared to traditional federated learning, both the privacy preservation and performance of SplitAD are improved. Through extensive experiments on three real-world anomaly detection datasets, SplitAD achieves a detection accuracy comparable to centralized training, and its communication overhead has a 71.7 ∼ 214.0 × improvement over existing privacy-preserving schemes. • A novel federated anomaly detection framework. • High-accuracy anomaly detection on vertically partitioned data. • Strong privacy preservation for both data and models. • Quite low system overhead via reconstruction error transmission.