Federated Learning for Deep Anomaly Detection with Noisy and Heterogeneous Data

We consider the problem of unsupervised learning for deep anomaly detection (DAD) in a federated learning (FL) network, consisting of a central server and many distributed clients. The conventional FedAvg algorithm involves clients training local DAD models with their private datasets, and then uploading these models to the server for aggregation into a global model. In practical scenarios, this framework faces two major challenges: 1) unlabeled training datasets may contain unknown anomalies; 2) training datasets from different clients are typically non-independent and identically distributed (non-IID). To address these problems, we propose FedDAD, a robust FL framework for training DAD models with noisy and heterogeneous data. In FedDAD, a small public dataset at the server, containing only a few normal samples (e.g., one sample from each normal class), serves as a normal anchor in the latent space across all clients. This anchor significantly improves the accuracy of identifying unknown anomalies at clients, in the presence of data heterogeneity. Having identified anomalies, clients utilize contrastive learning to train local feature extractors that further enhance the separation between normal and abnormal data. Extensive experimental results demonstrate the uniform superiority of FedDAD over all FL baselines across various settings and datasets. Furthermore, the model trained with FedDAD even achieves comparable performance to the model trained in a centralized manner.