Verifiable Federated Learning Based on Data Service Quality

In traditional distributed system, the central server has absolute management authority over data, while in federated learning, data is provided by task execution nodes, and the central node has no direct viewing authority. Task execution nodes provide data for local model training in a distributed manner, and the global model plays the role of aggregation of these local models. The quality of the global model is affected by the local models provided by task execution nodes, which are generated by algorithms and data. To measure the impact of data on the model, the concept of data service quality is introduced, which is defined as the accuracy, completeness, and relevance of the data to the task. High-quality data provides more reliable model updates, while low-quality data can cause model degradation. At the same time, the task node data service quality can be divided into malicious and non-malicious. Non-malicious behavior may be caused by data collection errors, while malicious behavior such as data poisoning may destroy the global model. A method for evaluating data service quality, node contribution, and credibility is designed to select high-quality nodes to participate in federated learning tasks. A large number of experiments were conducted on the public data set CIFAR-10, which verified the feasibility and superiority of VFL-DSQM.