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.

22 can be exploited to carry out privacy attacks, such as membership inference or partial reconstruction of the original training data [1],[2]. In this context, the primary objective of the study is to develop a mechanism for the effective integration of two innovative concepts—Federated Learning and Differential Privacy—to simultaneously enable distributed model training without sharing raw data and to ensure the protection of model updates exchanged between clients and the server [3]. Achieving this goal requires not only a theoretical analysis of the underlying principles but also an experimental evaluation of the proposed mechanism to assess its effectiveness and practical applicability in modern cybersecurity and machine learning environments. Before developing the mechanism itself, it is necessary to establish a solid theoretical foundation. The initial stage of the research focuses on a detailed analysis of the Federated Learning architecture, which involves training models directly on users‘ devices. In this process, only model updates (gradients) are transmitted to the central server, where they are aggregated, for instance, using the FedAvg algorithm. In parallel, the mathematical framework of Differential Privacy—the primary method for quantitatively defining privacy guarantees—is examined. This approach is characterized by the parameters $(\epsilon, \delta)$ and ensures that the presence or absence of a particular record in the training dataset has a minimal impact on the computation outcome. Another essential component of the theoretical analysis is the classification of contemporary privacy threats in Federated Learning, ranging from Membership Inference Attacks (MIA) to more advanced data reconstruction attacks that enable the recovery of original data solely based on model gradients [4]. The core part of the study is devoted to designing a model update algorithm that incorporates Differential Privacy within the Federated Learning environment. On the client side, this process is implemented as a two-stage method. The first stage involves gradient clipping, which limits gradient sensitivity and creates the necessary conditions for the correct application of privacy mechanisms. The second stage introduces the Gaussian mechanism, which adds random noise to the gradients. The