Federated Learning for Privacy-Preserving User Analytics: A Differentially Private, Heterogeneity-Aware Framework

User analytics often rely on sensitive behavioural data, which poses a privacy risk to individuals and a regulatory and security risk to the organisations that process it.Federated learning (FL) addresses this challenge by keeping data on the client devices where it is produced and only sharing model updates that contribute to the collective learning process, thus embedding the privacy principle of data minimisation into its architecture.However, FL is not without its privacy challenges, among them the potential for leakage through model updates, statistical heterogeneity of user data (Non-IID) impeding learning, and the costs (in terms of communication overhead and accuracy loss) of posttraining privacy guarantees.This paper investigates FL as a privacy-preserving solution for user analytics and introduces a framework that combines (i) improved client-side differential privacy (DP) mechanisms that utilise gradient clipping and add calibrated Gaussian noise to achieve (ε, δ)-DP, (ii) secure aggregation to hide the updates' true values from the server, (iii) a proximal term that improves convergence of local epochs in the presence of Non-IID data, and (iv) a privacy-budget allocation controller that balances the costs of differential privacy and communication.The framework follows the reference architecture described in IEEE 3652.1-2020.We formalise the problem, introduce the algorithm and threat model, and present an experimental protocol evaluating the framework's utility against centralised training and FedAvg.This paper reports the design and evaluation methodology; the metrics are currently being collected for the experimental phase.