Introduction: Software estimation plays a crucial role in predicting the budget and duration of software projects, and thus in their success. To estimate software, historical data are generally used, and machine learning is leveraged to predict software estimation. Vendors and companies share their past historical estimations with a central server, where a model is trained to evaluate and predict the estimation. However, all companies maintain confidential past estimations on their local servers. Objective: To eliminate this challenge, we use a federated learning framework to leverage AI and keep vendors' past estimates private in their local servers without sharing them with the central server. We also use machine learning to predict the software cost and duration. Material/Methods: In this framework, we train and evaluate the model in local servers and share their encrypted weights, bias, and accuracy score with a central server while ensuring past estimations are kept private. The central server will aggregate and average the weights and bias and share them back with the local server for retraining. This process will continue until it reaches the accuracy score necessary to predict and share the size of the new requirement with the central server and return the cost and duration. In the central server, we use a federated averaging algorithm for model aggregation, where the global model is updated by averaging the local model updates.

Result: The experimental results demonstrate the effectiveness of the proposed approach in accurate cost estimation and duration prediction by FLML. To evaluate the proposed framework, we conduct experiments using sample data. We compare the performance of the FL-based models with centralized models in terms of evaluation metrics. The performance will need to be improved (because of distributed/parallel data processing), but the FL framework provides privacy guarantees.