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

Area of Science:

• Software Engineering
• Machine Learning
• Data Privacy

Background:

• Software defect prediction (SDP) is crucial for identifying faulty modules early.
• Cross-project defect prediction (CPDP) faces challenges with data heterogeneity and privacy concerns.
• Existing methods often struggle with computational and communication overhead in federated environments.

Purpose of the Study:

• To introduce the Efficient Communication Federated Meta-Learning (ECFML) framework for privacy-preserving CPDP.
• To develop a lightweight yet effective model for software metrics representation learning.
• To evaluate ECFML's performance, privacy guarantees, and efficiency in federated settings.

Main Methods:

• ECFML integrates Model-Agnostic Meta-Learning (MAML) with a Mobile Vision Transformer (MobileViT)-inspired backbone for tabular software metrics.
• Feature vectors are transformed into token sequences processed by 1D convolutions and transformer layers.
• Laplace Differential Privacy is employed to ensure privacy guarantees in the federated learning process.

Main Results:

• ECFML demonstrates competitive or superior performance against ResNet-18 and U-Net on the AEEEM benchmark (EQ, JDT, PDE).
• The framework achieves significant gains in F1-score and AUC on the EQ dataset and consistent improvements on PDE.
• ECFML shows stable generalization across heterogeneous projects and maintains performance under certain noise and adversarial perturbations.

Conclusions:

• ECFML offers an effective balance between predictive accuracy, privacy preservation, and communication efficiency for federated SDP.
• The compact model design (∼142k parameters, ∼0.54MB) reduces computational and communication overhead.
• ECFML presents a viable solution for privacy-sensitive software repositories requiring robust defect prediction.