Authors: Kusum Kumari, Goutam Shaw, Debosmita Sukul, Anurima Majumdar, Antara Ghosal, Koushik Pal
Abstract: This paper discusses the advancement, challenges, and future of machine unlearning with emphasis on its significance in enhancing data privacy, security, and compliance with regulatory requirements. The review process began in 2015 and is ongoing to the current year. As privacy has become the focal point within the machine learning community, along with regulations like the General Data Protection Regulation (GDPR), machine unlearning—removing specific data from machine learning models—has attracted significant attention. The process of deleting such data is naturally timeconsuming, considering that it requires a complete retraining of the entire model; hence, traditional models have a dilemma because the process of erasing data is technically challenging and usually impractical considering the associated costs of computation. Machine unlearning enhances data privacy by facilitating selective erasure of specific data points without the need for total model retraining. It also improves model responsiveness and compliance with regulations like GDPR, hence encouraging the ethical application of artificial intelligence. The advantages of machine unlearning are enhanced data privacy, enhanced model performance, efficient utilization of resources, reduction of bias, quicker updates, and ensured compliance with ethics and laws. Through out the extensive literature survey a significant gap is observed to be that there are no reproducible, standardized procedures confirming the complete and effective elimination of data without compromising model efficiency and scalability. Areas of latent application in sectors like healthcare, finance, personalized services, and federated learning are identified, particularly in situations where unlearning is required to ensure privacy and compliance with regulations.
DOI: DOI: http://doi.org/10.5281/zenodo.15783199
