GAFSOM: A Privacy-Preserving Machine Learning Algorithm for SOM Clustering

Abstract Privacy-Preserving Data Mining (PPDM) requires methods that safeguard sensitive information while retaining the analytical value of data. This paper presents GAFSOM (Genetic-Fuzzy Algorithm for Self-Organizing Maps), a meta-heuristic anonymization framework that combines fuzzy sets with genetic optimization to balance privacy protection and clustering fidelity. By selectively applying fuzzification to high-risk attributes and leveraging genetic search to minimize distortion, GAFSOM preserves the topological structure of Self-Organizing Maps, an aspect often overlooked in existing anonymization techniques. The approach is evaluated on two benchmark datasets, UCI Adult and Bank Marketing, against a range of baselines including traditional SOM, k-anonymity, Fuzzy C-Means, genetic clustering, and differential privacy-enhanced SOM. Experimental results demonstrate that GAFSOM achieves superior clustering accuracy, lower information loss, and reduced topographic error, while maintaining competitive computational efficiency. Moreover, structural analyses confirm that the method preserves SOM’s neighborhood relationships with minimal distortion even under anonymization. These findings highlight GAFSOM as an effective and scalable solution for privacy-preserving, topology-sensitive data mining tasks.