Abnormal event detection in videos (VAD) serves as a key component in applications like intelligent surveillance and video content review, particularly in the context of growing concerns over video sequences privacy. Traditional anomaly detection methods typically require centralized training data, which poses privacy risks when dealing with sensitive data, such as surveillance videos or traffic accident footage. In practical scenarios, video sequences are often distributed across multiple institutions or devices, and due to privacy protection regulations, these videos cannot be shared. Therefore, how to efficiently per- form abnormal event detection in videos while ensuring privacy still constitutes a formidable challenge. To alleviate this issue, we propose a novel framework Federated Weakly Supervised Abnormal event detection in videos (Fed-WSVAD), which leverages federated learning to solve the problem of data privacy between institutions. Specifically, we combine the massive visual-textual pre-trained model CLIP with federated learning, introducing a global-level and local context-driven dynamic text prompt generator. This generator creates text prompts that enhance global-level generalization while maintaining local personalization dependent on the unique data characteristics of each client, consequently improving the effectiveness of abnormal event detection in videos. To further enhance the generalization ability of the model, we incorporate a memory module that stores and updates feature prototypes for normal and anomalous samples. This allows knowledge sharing and transfer across clients without requiring data sharing. Our model not only reduces data transmission and storage risks due to privacy concerns but also improves anomaly detection performance through the introduction of compactness and separateness losses. We implemented comprehensive experiments on the XD-Violence dataset and UCF-Crime dataset datasets, and the results show that compared to traditional federated learning methods, our Fed-WSVAD approach significantly outperforms in both global-level generalization and local personalization. This method effectively balances privacy protection and performance optimization, enabling the training of efficient abnormal event detection in videos models without sharing sensitive data. This research provides a novel solution for privacy-preserving abnormal event detection in videos and demonstrates the potential and practical application value of federated learning in weakly supervised abnormal event detection in videos. In the future, we will further explore how to optimize this method in more complex scenarios to enhance its feasibility and robustness in practical-scenario applications.
