Authors: Garibe Anuradha Gangadhar, Professor Abhishek Popat Bangar, Professor S. B. Bhosale, Professor Karad V. A.
Abstract: Trained machine learning models can hold on to traces of the data they were built from, including personal, outdated, incorrect, or unauthorized records. Deleting an entry from a database does not erase what a model has already absorbed from it. Machine unlearning tackles this gap by lowering or removing the effect of chosen training data, ideally without retraining the whole model every time. This review looks at five studies published between 2021 and 2025: one on coded machine unlearning, one on neuron masking, one on zero-shot unlearning, and two wider surveys of methods and evaluation. Together they trace a path from partition-based training designs and changes to model structure toward techniques that still work after the original training samples are gone, accompanied by increasingly systematic taxonomies and metrics. We compare the studies on the data access they assume, the unlearning strategy they use, their efficiency, how well they preserve utility, and how forgetting is verified. We also discuss the main open problems: the absence of a universal evaluation standard, the tension between forgetting and keeping performance on retained tasks, the risk of leftover information, and the difficulty of carrying these methods over to large language models and live systems. The work is a focused narrative review of the chosen papers; no new experiments or benchmark results are claimed. Our conclusion is that machine unlearning is a promising tool for privacy and model governance, but dependable use in practice requires measurable forgetting, clearly stated assumptions, and independent verification.
