A Critical Technical Analysis of Adaptive Privacy-budget Allocation in Federated Medical Imaging

This paper presents a critical technical analysis of adaptive privacy-budget allocation in federated medical imaging, supported by a structured
review of approaches published between 2015 and 2025. Twenty-two studies are examined, including seven addressing non-uniform allocation or related alternatives, focusing on allocation basis, protection granularity, conserved privacy quantity, and accounting compatibility. A reported allocation rule is analyzed using simulated radient-inversion similarity. Under the reported formulation, normalization by client size reduces the difference between the largest and smallest budgets, with the budget ratio approaching 1 and reaching approximately 1.0004 at a client size of 2172, rather than the intended eβ. A rank-normalized variant shows that preserving the sum of per-sample budgets does not preserve Rényi privacy cost, producing excess costs of
2.1%, 8.2%, and 31.3% for β = 0.5, 1, and 2, respectively. Using a reference accounting configuration constructed from available information, the least-protected record yields ϵ = 3.24 for a nominal ϵ = 2 at β = 1. The review identifies limitations in attack-derived sensitivity validation, privacy evaluation, and reporting of parameters required for reproducibility. Overall, the findings demonstrate that adaptive privacy allocation must be evaluated jointly with the conserved privacy quantity and the accountant used to establish privacy guarantees.