联邦学习环境下异构数据集的软件缺陷预测性能提升策略——以通信效率和隐私保护为中心的实证研究

Software defect prediction performance improvement strategy for heterogeneous datasets in a federated learning environment:empirical study centered on communication efficiency and privacy protection

为了解决联邦学习环境下异构数据集的软件缺陷预测中存在的通信效率低下和隐私保护问题,该研究利用Dueling双深Q网络(Double Deep Q-Network,DQN)算法的特性设计了一种软件缺陷预测算法,并通过构建模型来测试其性能;采用高斯差分隐私加密策略;针对通信效率问题,基于K-means模型进行聚焦;通过实验,文章针对不同数量的参与方以及通信轮次进行了详细的实验分析。研究发现,Dueling DQN算法在软件缺陷预测中表现出较高的准确性。高斯差分隐私加密策略在保护参与方数据隐私的同时,保持了模型的预测性能。K-means模型聚合策略在提升通信效率方面表现出显著优势。得出结论,通过采用Dueling DQN算法、差分隐私加密策略以及K-means模型聚合策略,联邦学习环境下异构数据集的软件缺陷预测性能得到了显著提升。

In order to address the problems of low communication inefficiency and privacy protection in the software defect prediction of heterogeneous data sets in federal learning environment,a software defect prediction algorithm is designed with the characteristics of Dueling DQN(Double Deep Q-Network)algorithm and testing the performance by building models;adopting Gaussian differential privacy encryption strategy;focusing on communication efficiency problems based on K-means model,and conducting detailed experimental analysis for different numbers of participants and communication rounds.We found that the Dueling DQN algorithm showed high accuracy in software defect prediction.The Gaussian differential privacy encryption strategy maintains the predictive performance of the model while protecting the data privacy of the participants.The K-means model aggregation strategy shows significant advantages in improving communication efficiency.It is concluded that the software defect prediction performance of heterogeneous data sets in the federated learning environment is significantly improved by adopting Dueling DQN algorithm,differential privacy encryption strategy and K-means model aggregation strategy.