Federated Learning for Video Action Recognition on UCF۱۰۱

With the proliferation of edge devices and growing concerns over data privacy, federated learning (FL) has emerged as a powerful paradigm for training machine learning models without centralizing raw data. In this paper, we apply FL to the task of video action recognition using the UCF۱۰۱ dataset. We simulate a realistic federated setting with multiple clients, each holding a local subset of videos, and train a ۳D ResNet model via the Federated Averaging (FedAvg) algorithm. We compare the performance of the federated model against a centralized baseline under both IID and non-IID data partitions. Under the IID partition, the federated model achieves ۸۳.۱% test accuracy-only ۱.۶ percentage points below the centralized baseline-while under a challenging pathological non-IID partition, it reaches ۷۶.۴%. We also analyze the effect of local epochs on convergence and demonstrate significant communication savings. Our work provides a practical framework for deploying privacy-preserving video analytics at the edge and highlights the challenges of heterogeneous client data.

کلیدواژه ها:

Federated Learning ، Video Action Recognition ، UCF۱۰۱ ، FedAvg ، Privacy-Preserving Machine Learning ، Edge Computing