Adaptive Weighted Federated Learning for Enhanced Data Privacy in Cross-Platform E-commerce

Abstract

Aiming at the urgent need for data privacy protection of cross-border e-commerce platforms, this project proposes a privacy-enhanced federated learning algorithm based on weight fusion. The algorithm can dynamically adjust the aggregation weight of parameters according to the data volume, quality, and feature distribution of participants, and achieve a balance between privacy protection and model performance. In terms of experiments, a simulated cross-platform e-commerce database was established and compared with algorithms such as FedAvg and FedDP. The experimental results show that after 100 iterations, the accuracy of the test case set of the algorithm reaches 85.3%, which is 7.2 percentage points higher than FedAvg's 78.1%; in terms of convergence speed, the algorithm proposed in this paper converges in 52 iterations, while FedAvg needs 75 times and FedDP needs 70 times; in terms of privacy protection, under the same privacy budget conditions, the algorithm in this paper can reduce the success rate of model inversion attacks to 4.8% and 5.2% respectively, which is much lower than FedDP's 12.3% and 11.8%. Adaptive weights dynamically adjust via multi-dimensional evaluation (data scale: 40%, quality: 30%, features: 30%), while hierarchical privacy allocates ε=0.1 for sensitive data (financial) and ε=0.5 for non-sensitive (clicks), boosting accuracy by 7.2% and cutting attacks to 4.8%. The experimental results show that the algorithm proposed in this paper can effectively improve the data privacy protection level and model training effect of cross-platform e-commerce platforms, and its performance is significantly better than SOTA federated learning algorithms. This work pushes the development of federated learning in e-commerce data privacy protection by proposing a new framework of adaptive weighted aggregation combined with hierarchical differential privacy, which provides a theoretical and experimental basis for cross-platform data collaboration under strict data privacy regulations.