Distributed Anomaly Detection Using Federated Learning

One of the major challenges video surveillance systems usually face is high communication costs and computational bottlenecks, leading to latency in the transfer of high-definition video data. Such issues generally arise when large-scale projects are implemented in real-time capabilities. The paper explores Federated Learning (FL) algorithms for an efficient and scalable suspicious activity detection system to address these limitations. Federated Learning allows a decentralized approach that allows model training on participating devices (clients), and only the model updates are shared with the central server instead of sharing a huge amount of data. The approach reduces bandwidth consumption and minimizes latency without compromising the performance of the model. This study is among the first to explore federated learning (FL) for video-based anomaly detection using the UCF-Crime dataset. We propose and comparatively evaluate three FL aggregation strategies—FedAvg, Unweighted FedAvg, and FedSGD—under realistic, non-IID client data distributions. Unlike existing centralized methods, our approach offers enhanced privacy and reduced communication overhead while maintaining competitive performance. The results obtained in the experiments conducted on a varying number of clients for each approach suggest Unweighted FedAvg as the most reliable approach, which consistently achieves the highest accuracy and lowest loss for any given number of clients. The Accuracy for FedAvg ranges from 0.73 for 5 clients to 0.82 for 30 clients, and the loss decreases from 0.81 for 5 clients to 0.64 for 30 clients. Similarly, Unweighted FedAvg shows a consistent increase in accuracy from 0.74 for 5 clients to 0.85 for 25 clients and then drops slightly to 0.84 for 30 clients, and similarly for loss also it decreases from 0.78 for 5 clients to 0.53 for 25 clients and increases slightly to 0.56 for 30 clients. Despite the slight decrease in accuracy and increase in loss, it always performs better than FeDAvg. Whereas FedSGD has shown a very erratic pattern in change of accuracy and loss as the number of clients increases, with the lowest accuracy being 0.29 and the highest being 0.41, and for loss highest was 1.91 and the lowest was 1.59. The study successfully uncovers the potential of federated learning to address the limitations, such as computational bottlenecks, latency in data transfer, and model performance over non-iid data in real-time implementations of the public security systems.