本地差分隐私下基于聚类的两阶段多任务学习算法OA

The widespread application of multi-task learning in data analysis amplifies privacy leakage risks.Traditional approaches relying on trusted servers,neglecting noise interference,and considering only task correlation result in insuffi-cient privacy protection and degraded model performance.This study proposes KTMTL,a two-stage clustering-based multi-task learning algorithm under local differential privacy(LDP),to jointly optimize privacy preservation and model utility.The first stage replaces the Euclidean distance with Huber distance in an improved K-means clustering algorithm,effectively suppressing outlier impacts induced by Laplace noise.The second stage constructs an interactive multi-task learning framework that simultaneously models feature-task correlations and optimizes parameters through gradient aggre-gation.Theoretical analysis confirms the algorithm's rigorous LDP compliance and bounded complexity.Experiments on School,ADNI,and synthetic datasets demonstrate KTMTL outperforms existing privacy-preserving algorithms(e.g.,DP-MTRL,DP-DMTL)by 10%-15%in prediction accuracy(AUC)and reduces normalized mean squared error(nMSE)by 8%-12%under equivalent LDP guarantees,while significantly outperforming comparable methods in computational efficiency.The proposed framework provides an efficient and robust solution for privacy-sensitive multi-task learning scenarios through synergistic optimization of privacy protection and task grouping quality.