A feature-perturbed local label differential privacy algorithm

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This paper investigates the privacy-utility trade-off of local label differential privacy (L-LDP) in deep learning tasks such as image classification. Theoretical analysis reveals that the privacy enhancement of L-LDP increases linearly with the privacy budget, while model utility degrades in an inverse-proportional manner; this intrinsic imbalance makes it difficult for L-LDP to maintain a stable trade-off across different privacy levels. To address this limitation, an improved L-LDP algorithm based on feature perturbation and structured resampling is proposed. The proposed method preserves the original privacy guarantees of L-LDP while enhancing data diversity and model robustness through feature-space clustering and randomized resampling. Experimental results demonstrate that the proposed algorithm significantly improves privacy protection without sacrificing model utility, achieving a more favorable balance between privacy and performance across multiple image classification tasks.