Privacy-Utility-Efficiency Trade-Offs in Personalized Federated Learning for Edge Computing

For the edge applications of Industrial Internet of Things (IIoT), the task of learning from distributed and privacy-conscious data needs to be conducted under constrained communication resources and in the presence of highly heterogeneous clients. This paper offers a personalized Federated Learning (FL) solution where the global backbone can be decoupled from the per-client heads, enabling the concurrent execution of global representation learning and local adaptation, even with non-Independent and Identically Distributed (non-IID) data. Differential Privacy (DP) is used to ensure privacy protection, where the updates of the backbone are differentially privatized via clipping and Gaussian noise, with the total privacy budget (ε, δ) measured over the number of communication rounds via a Rényi Differential Privacy (RDP) accountant. This study also provides a set of training diagnostics related to clipping pressure, including the proportion of clipped clients and the dynamics of the norms of the updates, which can identify the points at which privacy noise begins to dominate the learning process, along with privacy-friendly operating configurations. Experimental results on the LEAF and FEMNIST datasets reveal higher average accuracy and a narrower per-client accuracy distribution than those of the FL counterparts, whereas the privacy–utility–efficiency analysis identifies Pareto points at which privacy can be improved with a constant per-round communication cost, since only the backbone updates are communicated. In practice, this means more reliable performance at the client level for controlled privacy loss and predictable communication overheads, which makes this approach appropriate for edge sites with varying quality and availability of data.