Software Defect Prediction (SDP) guarantees software reliability by identifying defect modules before release. The existing machine learning models, however, rely on centralized collection of data, thus raising concern over privacy, security, and scalability of the process. The paper thus proposes a federated learning (FL) based decentralized approach to software defect prediction using FedProx. The proposed scheme establishes the setting for the joint model training over several clients without exposing sensitive information regarding software repositories. Our approach symbiotically integrates FedProx-a federated optimization method that counteracts system heterogeneity and enhances model convergence. The implementation is on the NASA PROMISE datasets (CM1, KC1, PC1) using neural networks in defect prediction. Further improvements to model performance include ADASYN-based class balancing, polynomial feature engineering, and weighted loss functions. Experimental results show that our FedProx-guided FL methodology exhibits better defect prediction accuracy while maintaining privacy and computational efficiency. The comparison shows its advantages against traditional centralized models and individual ones. These findings reveal that federated learning potentially offers real-world software defect prediction as a scalable and yet privacy-preserving solution for software development in the future. Experiments conducted on the CM1, KC1, and PC1 datasets achieved F1-scores of 84.22%, 93.84%, and 95.94%, respectively, along with recall values exceeding 96%, and consistently high accuracy and precision across all three datasets. The results achieved confirmed the effectiveness of our approach.