Heterogeneous defect prediction based on robust federated learning

In the realm of software engineering, accurate software defect prediction (SDP) is paramount for improving quality assurance processes. Traditional SDP methods face challenges such as data privacy concerns, heterogeneity, and label noise, especially in cross-project scenarios. To address these issues, this paper introduces the Heterogeneous Defect Prediction Based on Robust Federated Learning (HRFL). This novel framework utilizes federated learning to enable collaborative training across diverse data sources while preserving data privacy. HRFL integrates One-Side Selection (OSS) and Bootstrap Loss to enhance robustness against noisy labels and adjusts the influence of each participant with a Client Confidence Reweighting (CCR) mechanism based on data quality. The framework's efficacy is demonstrated through extensive experiments on three public datasets—NASA, AEEEM, and Relink—showing notable improvements in defect prediction accuracy under high noise conditions and data diversity. These results affirm HRFL's potential to significantly advance SDP practices by leveraging federated learning for enhanced privacy, scalability, and effectiveness. This study underscores the adaptability of federated learning in addressing complex challenges in SDP and other domains where privacy and data heterogeneity are paramount.