Federated model with contrastive learning and adaptive control variates for human activity recognition

• FedCoad integrates contrastive learning with adaptive control variates to mitigate non-IID data skewness in federated human activity recognition. • The proposed method reduces representation gaps between global and local models, improving convergence despite heterogeneous client data. • Adaptive control variates penalize local updates based on weight magnitude and update rate, effectively preventing objective drift. • FedCoad outperforms state-of-the-art federated learning algorithms on HAR benchmark datasets, offering a robust privacy-preserving solution.