基于方波机制的差分隐私域自适应学习方法OA

Domain adaptation became an increasingly important technique for addressing distributional shifts and label scarcity across domains.However,existing privacy-preserving methods,such as differential privacy,often suffered from excessive noise injection and limited model performance,especially in untrusted and high-dimensional settings.To address these challenges,this paper proposed a local differentially private square wave-based domain adaptation method called LDP-SWDA.Specifically,LDP-SWDA applied square wave-based local perturbation to the feature covariance structure with high probability.A gradient-based correction method was employed to restore the positive semi-definiteness of the covariance matrix,which might have been compromised by perturbation.Theoretical analysis was conducted on the privacy of LDP-SWDA,and the effectiveness of LDP-SWDA was evaluated on two standard domain adaptive learning datasets.The results indicated that LDP-SWDA exhibited good practicality and robustness in high-dimensional and privacy-sensitive scenarios,which had research significance.