Proceedings of the International scientific and practical conference ―Science, Technology and Culture: Dynamics of Change in the XXI Century‖ (December 1921, 2025) / Publisher website: www.naukainfo.com. – Baku, Azerbaijan, 2026. – 90 p.

23 magnitude of this noise is determined by the predefined privacy budget $\epsilon$ and the clipping threshold $C$. On the server side, the collected noisy updates from all participants are aggregated, for example, using the FedAvg method. As a result, a new global model is produced that guarantees a certain level of privacy protection. The analysis also justifies the choice of a specific machine learning model (in particular, a convolutional neural network for classification tasks) and determines the optimal values of the parameters $\epsilon$ and $C$. This ensures a balanced trade-off between model accuracy and the degree of data protection. The practical value of the research is validated through a series of detailed experiments. These experiments involve the creation of a test environment to simulate the distributed learning process using a selected dataset, such as CIFAR-10, partitioned among 100 clients. A central focus of the experimental evaluation is the analysis of the trade-off between model accuracy and privacy guarantees. Specifically, the study examines how the performance of the final model changes as the value of the parameter $\epsilon$ is gradually reduced, thereby strengthening privacy guarantees. The final and most critical stage involves assessing the proposed approach‘s resistance to data leakage attacks [5]. To this end, the FL-DP method is tested against a Membership Inference Attack applied to a model trained using the proposed mechanism. The results are compared with those obtained from a similar attack on a standard Federated Learning model without privacy protection. Privacy is considered strong if the MIA accuracy approaches random guessing, that is, approximately 50%. The final results are presented in the form of tables and graphs, clearly demonstrating the advantages and effectiveness of the proposed mechanism compared to baseline approaches. In the concluding section, the obtained results are summarized, and the degree to which the stated objective has been achieved is assessed. The main emphasis is placed on conclusions regarding the effectiveness of the proposed mechanism, particularly its ability to provide mathematically proven privacy guarantees without compromising acceptable—and in some cases high—model accuracy. The significant