Noise Placement, Privacy Accounting, and Structured Clipping in Client-Level Differentially Private Federated Learning An Empirical Study

This paper presents a controlled empirical study of three implementation choices in client-level differentially private federated learning: noise placement, privacy accounting, and structured clipping. Experiments are conducted under a trusted-server threat model with client-level add/remove adjacency across CIFAR-10, Adult, and synthetic datasets. The study compares central and distributed noise placement, classical composition and zCDP accounting, and semantic versus global clipping. Results show measurable differences in utility across privacy mechanisms and accounting choices, while semantic clipping does not consistently improve accuracy or minority recall. The study does not introduce a new differential privacy mechanism and provides code, configurations, seeds, and result files for reproducibility.