FedVAR: A Federated Learning Framework for Video Anomaly Recognition
A new framework called FedVAR has been developed by researchers to tackle the issue of semantic misalignment in Video Anomaly Recognition (VAR) among distributed edge clients. This framework is particularly significant for Industrial Internet of Things (IIoT) and Cyber-Physical Systems (CPS), where it is crucial to maintain high-fidelity Digital Twins and ensure safety in critical environments. FedVAR addresses the challenge of data heterogeneity among clients, which results in differing feature representations for normal and abnormal events, especially with fine-grained anomaly categories. Current federated methods primarily focus on binary anomaly detection and do not resolve this misalignment, hindering fine-grained recognition. By utilizing a weakly-supervised method to align prototypes across clients, FedVAR enhances anomaly recognition accuracy. The paper can be found on arXiv with the identifier 2608.06876, categorized as a cross announcement.
Key facts
- FedVAR is a prototype-aligned federated framework for Video Anomaly Recognition (VAR).
- It addresses semantic misalignment caused by data heterogeneity across distributed edge clients.
- The framework is designed for IIoT and CPS environments, supporting Digital Twins and safety.
- Existing federated methods focus on binary anomaly detection and fail on fine-grained recognition.
- FedVAR uses weakly-supervised learning to align prototypes across clients.
- The paper is available on arXiv with identifier 2608.06876.
- The announcement type is cross.
- The framework aims to improve fine-grained anomaly recognition in federated settings.
Entities
Institutions
- arXiv