A privacy-preserving federated meta-learning framework for cross-project defect prediction in software systems

Software defect prediction (SDP) is a critical task in software engineering, aiming to identify fault-prone modules before deployment. This paper introduces the Efficient Communication Federated Meta-Learning (ECFML) framework for cross-project defect prediction (CPDP). ECFML integrates Model-Agnostic Meta-Learning (MAML) with a lightweight Mobile Vision Transformer (MobileViT)-inspired backbone adapted for tabular software metrics. Feature vectors are projected into token sequences and processed via 1D convolutions and transformer mixing, enabling effective representation learning with a compact footprint (∼142 k parameters, ∼0.54 MB). This design reduces both computation and communication overhead in federated environments. Experiments on the AEEEM benchmark (EQ, JDT, PDE) show that ECFML achieves competitive or superior performance compared to ResNet-18 and U-Net. On EQ, it yields the highest gains in F1-score and AUC; on PDE it consistently improves F1-score and G-Mean; and on JDT it achieves performance comparable to strong baselines, reflecting stable generalization across heterogeneous projects. Privacy is enforced via Laplace Differential Privacy with a fixed clipping bound specified a priori, ensuring pure ɛ-DP guarantees per round under conservative composition ([Formula: see text]). Robustness analysis further shows that the framework maintains stronger performance than baselines under additive Gaussian noise and FGSM perturbations, though degradation remains under stronger adversarial settings. Overall, ECFML strikes a balance between predictive accuracy, privacy preservation, and communication efficiency, making it a viable solution for federated, privacy-sensitive software repositories.