Local Adaptive Differential Privacy for Federated Learning With Entity‐Aware Text and Feature‐Wise Clipping
TL;DR
Local Adaptive Differential Privacy (LADP), a client‐side privacy framework that allocates bounded privacy budgets, provides a practical, modality‐unified privacy layer for heterogeneous FL conditions.
Abstract
Federated learning (FL) keeps raw data on client devices, but model updates and learned representations can still leak sensitive information. We propose Local Adaptive Differential Privacy (LADP), a client‐side privacy framework that allocates bounded privacy budgets, ε∈[0.05,0.5]$$ \varepsilon \in \left[\mathrm{0.05,0.5}\right] $$ , at fine granularity: per feature for tabular data and per entity for text using Named Entity Recognition (NER). LADP computes budgets from lightweight local risk signals, applies feature‐wise clipping and Laplace perturbation for tabular attributes, and uses randomized response and the exponential mechanism for entity‐aware text perturbation. Experiments on healthcare tabular task and multiple text corpora under centralized and federated learning, including heterogeneous client partitions, show a controllable privacy–utility trade‐off. At moderate privacy budgets, LADP preserves predictive performance with only marginal degradation while substantially reducing membership inference and entity/attribute re‐identification risks. Across settings, Random Forest (RF) remains the most robust tabular model, and results are stable under outlier handling. Collectively, LADP provides a practical, modality‐unified privacy layer for heterogeneous FL conditions.