Real-Time Dynamic Multimodal Federated Anomaly Detection With Explainable Temporal Graph Learning

View at publisher Summarise with VIRAT
Authors: Nilamadhab Mishra, Vanamala Vennela, V Meghana Reddy, Chikkam Sravanth Royal
Cited by: 0
References: 24

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The dynamic evolutionary nature of fraud practices in the contemporary e-commerce systems has necessitated real-time, dynamically adaptive, scalable, and privacy-sensitive anomaly detection systems. Although the currently available GNN-based models exhibit a better detection performance, most of them support only a static graph structure, centralized training, and unimodal data processing, which limits their use in large and dynamically changing markets. This paper attempts to counter these shortcomings by introducing RT-FedXGAD, which is a Real-Time Dynamic Multimodal Federated Anomaly Detection system that is also provided with Explainable Temporal Graph Learning. The suggested architecture process includes multimodal data, transaction metadata, review text embeddings, behavioral features, and visual product cues in the temporal heterogeneous graph representation, which is updated according to online streaming activities. The federated aggregation mechanism is based on the idea of decentralization of the model training on distributed e-commerce systems, without the transfer of sensitive user information, which guarantees compliance with regulatory requirements under the GDPR and privacy guidelines. There is an explainable temporal graph attention module (XT-GAT) that is added to offer interpretable anomaly scores, which points out influential nodes, time windows, and modality contributions. Initial analyses of simulated real-time streams based on Amazon and Alibaba data show that they are much better at performance and achieve lower detection latency, better AUC-ROC, and more robustness to unknown fraud patterns than centralized static GNN baselines. The research is a contribution to the deployable, privacy-conscious, real-time fraud detection systems of large-scale commercial systems.

Fields of Study & Subjects

Privacy-Preserving Technologies in Data Anomaly Detection Techniques and Applications Advanced Graph Neural Networks

Publication details

Journal / source
International Conference on Communication, Computing and Electronics Systems (ICCCES)
Pages
1427–1434
Publication date
21 January 2026
Language
EN