A Novel Secure Approach for Enhancing Accuracy of Pest Detection with Private Federated Learning Using DPSGD

Federated learning is a technique that protects the privacy of data. Rather than sharing raw data, end users transmit their model parameters to a central aggregator or server. The central aggregator then combines all received parameters to form a final model, which is then distributed to clients. By adopting this approach, clients can avoid sharing their data with the server, instead sharing only model parameters. However, this approach poses a risk to data security, as attackers may be able to recreate the original dataset by analyzing the distributed model parameters. To address this issue, we propose a novel approach on agriculture pest dataset that utilizes differentially private parameters, which are sent to the central aggregator to improve data privacy. We conducted an analysis of the central model's accuracy by varying key model parameters, including epoch, learning rate, and iterations. Our findings indicate that maintaining moderate values for these parameters results in better accuracy for the proposed private federated learning model.