Enhancing Data Privacy Protection and Efficiency in Distributed Learning through Adaptive Communication Techniques

With the gradual proliferation of smart devices in recent years, distributed learning has become increasingly common and important in various practical scenarios. However, the time consumption during the communication process has also become a significant issue that cannot be overlooked. Nevertheless, there is still no sufficiently universal communication algorithm that avoids involving sensitive data while mitigating related privacy concerns and security risks. Therefore, this paper collects runtime data from different devices and utilizes a biased label predictor based on recursive simulation principles to create a dataset for pre-training various adaptive communication algorithms that dynamically adjust the number of local training rounds. The author will test the performance of this adaptive communication algorithm in optimizing the efficiency of a local update stochastic gradient descent (SGD) algorithm simulated to run on four different devices, as well as its adaptive effect on the local update rounds parameter. To enhance the robustness of this algorithm, the author also introduced an algorithm update mechanism and compared it with our pre-trained model. The results show that both the pre-trained communication algorithm and the updating algorithm can adjust local update rounds to some extent. While algorithm updates involve a trade-off between time efficiency and generalization, the dynamic model achieves superior computational efficiency and typically higher overall training efficiency compared to baseline methods, despite occasional overheads from dynamic mechanisms.