Enhancing Software Defect Prediction Security via Federated BLSTM with Differential Privacy

Mon, 22 Jun 2026 00:00:00 +0800

Software Defect Prediction (SDP) is an issue of paramount importance for improvement of software quality but data isolation and privacy concerns across organizations affects efficacy of SDP. By making use of a Federated Learning (FL) framework in combination with Differential Privacy (DP) for collective learning without exchanging any raw data, we provide a unique solution to this problem. The study makes use of advanced deep learning architectures the baseline Neural Network (NN), the sequential Long Short-Term Memory (LSTM) and the contextual Bidirectional LSTM (BLSTM) that were trained on the ten different, non-IID client datasets (software projects). Data processing included logarithmic transformation, Min Max Scaling, while the Synthetic Minority Over-sampling Technique (SMOTE) algorithm was used to address local class imbalance. The application of DP noise ensured the privacy of the model updates by quantifying them before aggregation using FedAvg. The results are a definite success for recurrent models with the BLSTM network coming on top, achieving a peak Global Accuracy of 99.05% and an F1-Score of 98.9% with the constraint of FL-DP. In comparison, the centralized, non-private BLSTM had slightly better performance in terms of accuracy, 99.3%, proving that the FL-DP framework is able to preserve close to optimal performance while increasing data security drastically. These results confirm that FL-DP is a powerful, safe and open-source paradigm for collaborative SDP, which enables a base solution for organizations that need to achieve high predictive power while maintaining high data confidentiality.

Journal of Electronic & Information Systems

Enhancing Software Defect Prediction Security via Federated BLSTM with Differential Privacy