Differential privacy (DP) provides a mathematically grounded framework for safeguarding sensitive information in large datasets; however, traditional clustering methods often suffer reduced accuracy under strict privacy constraints. This paper presents a local distance-based differential privacy model for self-organizing maps (SOM), designed to preserve data utility while ensuring strong privacy protection. By introducing noise proportional to local distances within the SOM’s neighborhood update process, the proposed method minimizes quantization error and maintains topological consistency. Unlike global DP methods such as DP k-means or conventional local DP models that perturb centroids independently, this approach achieves stable convergence, preserves structural integrity, and flexibly adjusts to varying privacy budgets ( ϵ ). Performance evaluation on several benchmark datasets emphasizes accuracy and topology preservation under DP constraints. Comparative analyses with state-of-the-art SOM clustering techniques highlight three major benefits: (1) preservation of neighborhood and topological structure, (2) stable convergence with low quantization error, and (3) scalability to high-dimensional data without loss of cluster coherence. Overall, the proposed framework demonstrates an improved privacy–utility trade-off compared to existing differentially private clustering methods.

Top