Abstract
Anomaly detection on attributed graphs has applications in various domains such as finance and email spam detection, thus gaining substantial attention. Distributed scenarios can also involve issues related to anomaly detection in attribute graphs, such as in medical scenarios. However, most of the existing anomaly detection methods are designed for centralized scenarios, and directly applying them to distributed settings may lead to reduced performance. One possible reason for this issue is that, when graph data are distributed across multiple clients, federated graph learning may struggle to fully exploit the potential of the dispersed data, leading to suboptimal performance. Building on this insight, we propose FedCLGN, a federated graph anomaly detection framework that leverages contrastive self-supervised learning. First, we put forward an augmentation method to maintain global negative pairs on the server. This involves identifying anomalous nodes using pseudo-labels, extracting embedding representations of the negative pairs corresponding to these anomalous nodes from clients, and uploading them to the server. Then, we adopt graph diffusion to enhance the feature representation of nodes, capturing the global structure and local connection patterns. This strategy can strengthen the differentiation between positive and negative instance pairs. Finally, the effectiveness of our approach is verified by experimental results on four real graph datasets.
Introduction
Anomaly detection on graphs—such as identifying fraudulent accounts in social networks or rare diseases in medical records—typically requires a holistic view of the entire network. However, in real-world applications, data is often siloed across different organizations due to privacy regulations. When applying Federated Learning (FL) to this problem, two major hurdles arise: Information Scarcity, where individual clients lack enough negative examples to train robust models, and Missing Links, where structural relationships crossing client boundaries are lost.
Existing methods often struggle to bridge this gap without leaking sensitive raw features. To address these challenges, we introduce FedCLGN (Federated Contrastive Learning with Global Negative pairs), a novel framework designed to detect anomalies in attributed graphs without compromising data privacy.
Methodology
FedCLGN utilizes a contrastive self-supervised learning approach consisting of four key components:
- Local Contrastive Learning: Each client trains a local model that compares a target node to its local subgraph. Nodes that do not "fit" their neighborhood are flagged as potential anomalies.
- Global Negative Pair Strategy: This is the core innovation. Clients use pseudo-labels to identify likely anomalous nodes. Instead of sharing raw data, they upload high-level embedding representations of these negative pairs to a central server, which aggregates them into a global pool.
- Graph Diffusion: To compensate for missing cross-client edges, the server applies graph diffusion (e.g., Personalized PageRank) to the global embedding pool. This captures broader structural patterns and strengthens the differentiation between positive and negative instance pairs.
- Personalized Parameter Aggregation: Unlike simple averaging, the server calculates similarity between clients based on global negative pairs to perform personalized weight updates.
Experimental Evaluation
We evaluated FedCLGN on four widely used graph benchmarks: citation networks (Cora and Citeseer) and social networks (BlogCatalog and Flickr). Since these datasets do not have natural anomalies, we followed standard practice by injecting synthetic anomalies—swapping node attributes with those of structurally distant nodes—to test detection capabilities.
Key Results
FedCLGN consistently outperformed several state-of-the-art baselines, including centralized models (DOMINANT, CoLA) and federated versions (FedAvg-CoLA).
- Superior Accuracy: On the Cora dataset with 5 clients, FedCLGN achieved an AUC of 0.861, significantly higher than the 0.803 achieved by FedAvg-CoLA.
- Scalability: The model maintained high performance even as data became more fragmented across 10 and 15 clients, whereas baselines saw sharper performance declines.
- Robustness: Sensitivity analysis demonstrated that the model remains stable across different confidence thresholds for pseudo-anomaly identification.
Conclusion
FedCLGN represents a significant advancement in federated graph anomaly detection. By sharing high-level embeddings of potential anomalies rather than raw features, the framework improves detection accuracy in data-sparse environments while maintaining privacy. Future work will focus on optimizing the framework for extreme non-IID scenarios and dynamic, evolving graphs.
Summary updated