A Novel User Behavior Modeling Scheme for Edge Devices with Dynamic Privacy Budget Allocation

Federated learning (FL) enables privacy-preserving collaborative model training across edge devices without exposing raw user data, but it is vulnerable to privacy leakage through shared model updates, making differential privacy (DP) essential. Existing DP-based FL methods, such as fixed-noise DP, suffer from excessive noise injection and inefficient privacy budget allocation, which degrade model accuracy. To address these limitations, we propose an adaptive differential privacy mechanism that dynamically adjusts the noise based on gradient sensitivity, optimizing the privacy–accuracy trade-off, along with a hierarchical privacy budget management strategy to minimize cumulative privacy loss. We also incorporate communication-efficient techniques like gradient sparsification and quantization to reduce bandwidth usage without sacrificing privacy guarantees. Experimental results on three real-world datasets showed that our adaptive DP-FL method improved accuracy by up to 8.1%, reduced privacy loss by 38%, and lowered communication overhead by 15–18%. While promising, our method’s robustness against advanced privacy attacks and its scalability in real-world edge environments are areas for future exploration, highlighting the need for further validation in practical FL applications such as personalized recommendation and privacy-sensitive user behavior modeling.