Key points are not available for this paper at this time.
Consumer electronic devices often involve processing and analyzing a large amount of user personal data. Nevertheless, owing to apprehensions regarding privacy and security, users are hesitant to transmit this sensitive data to centralized cloud servers for training. The combination of mobile edge computing and federated learning (FL) enables local devices to access computational power and storage resources, allowing them to engage in distributed learning and model training while safeguarding user privacy. However, these resources are not unlimited. Furthermore, as artificial intelligence technology progresses, inference attacks have become a major threat to privacy in traditional federated learning. To address these challenges, we propose an innovative federated deep learning algorithm, called Fed-PEMC. This algorithm combines local differential privacy and model compression techniques. By leveraging deep reinforcement learning for model compression, Fed-PEMC reduces model size while maintaining model accuracy, improving communication efficiency. We also introduce customized label sampling to accelerate model training. Before uploading the model, we implement local differential privacy protection on the compressed model, reducing privacy budget and addressing privacy leakage caused by inference attacks. Theoretical analysis and experimental results validate that Fed-PEMC adheres to (ϵ, δ)-differential privacy and exhibits a communication cost linked to the model size. Experimental results show that compared to baseline algorithms, Fed-PEMC excels in ensuring privacy, maintaining model accuracy, and optimizing communication efficiency, and Fed-PEMC outperforms the baseline solution DP-Fed by 2.27 and 2.02 percentage points in testing accuracy on the Mnist and Cifar10 datasets, respectively.
Lin et al. (Tue,) studied this question.