Machine learning–driven data perturbation techniques for privacy-preserving data mining

Description

Sunil and Narsimha build a privacy shield for data mining in 2026 that scrambles records before analysis instead of locking them away or stripping out names. Their framework clusters similar records first, then applies a flip-and-rotation perturbation, a controlled shuffle of values, chosen feature by feature with a structured decision method. Older anonymization tends to gut a dataset's usefulness, while encryption is heavy to compute, so perturbation aims at the middle ground.

Tested against a conventional Naive Bayes approach and fuzzy-based privacy methods, the new framework scored better on all four yardsticks the authors used for judging mining quality. Those yardsticks, accuracy, precision, recall, and the combined F-measure, each favored the perturbation design on the structured datasets used in the experiments. The authors conclude the balance is practical, keeping personal details protected while leaving the patterns analysts mine for still readable in the data.

Catalog record

Source: https://europepmc.org/articles/PMC13482305