Scholarlink

<p>This study presents a comprehensive literature review on Machine Unlearning (MU), which aims to eliminate the influence of specific data samples from trained machine learning models under the “right to be forgotten” framework mandated by the European Union GDPR and similar regulations. This study presents a systematic literature review of MU research published between 2020 and 2025. More than 170 publications were reviewed. A taxonomy of centralized and Federated Unlearning(FU) methods is proposed, along with four main categories (Exact, Approximate, Certified, and Verifiable Unlearning). Findings are summarized across benchmark datasets, including MNIST, CIFAR-10, CIFAR-100, ImageNet, IMDB, AG News, MovieLens, as well as healthcare and finance related tabular datasets. Across these benchmarks, prior studies consistently report that approximate unlearning methods offer substantial computational advantages over full retraining while maintaining competitive model utility, whereas certified methods provide stronger formal guarantees at the cost of increased computational overhead. However, these deployments present important security risks like membership inference, data poisoning, malicious forgetting requests, data re-engineering, and model inversion attacks, as well as significant engineering challenges related to energy efficiency, communication overhead, and scalability constraints in IoT systems. The study also reviews privacy-enhancing approaches that integrate differential privacy, homomorphic encryption, secure computation, trusted computing environments, and access control mechanisms into the unlearning process. The viability of regulatory compliance is analyzed from a sectoral viewpoint in terms of ethics, justice, and accountability, especially in high-security industries like healthcare and finance. These factors highlight the necessity of scalable, verifiable, attack-resistant, and resource-efficient MU techniques for privacy-preserving AI.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine unlearning through model forgetting techniques: application area and emerging methods

  • 期刊论文

  • 发表日期：

    2026年7月24日

查看全文
  • Busra Buyuktanir,
  • Elif Kucur,
  • Zeynep Kuyumcu,
  • Nese Ozdemir,
  • Sahsene Altinkaya,
  • Gozde Karatas Baydogmus,
  • Kazim Yildiz

摘要

This study presents a comprehensive literature review on Machine Unlearning (MU), which aims to eliminate the influence of specific data samples from trained machine learning models under the “right to be forgotten” framework mandated by the European Union GDPR and similar regulations. This study presents a systematic literature review of MU research published between 2020 and 2025. More than 170 publications were reviewed. A taxonomy of centralized and Federated Unlearning(FU) methods is proposed, along with four main categories (Exact, Approximate, Certified, and Verifiable Unlearning). Findings are summarized across benchmark datasets, including MNIST, CIFAR-10, CIFAR-100, ImageNet, IMDB, AG News, MovieLens, as well as healthcare and finance related tabular datasets. Across these benchmarks, prior studies consistently report that approximate unlearning methods offer substantial computational advantages over full retraining while maintaining competitive model utility, whereas certified methods provide stronger formal guarantees at the cost of increased computational overhead. However, these deployments present important security risks like membership inference, data poisoning, malicious forgetting requests, data re-engineering, and model inversion attacks, as well as significant engineering challenges related to energy efficiency, communication overhead, and scalability constraints in IoT systems. The study also reviews privacy-enhancing approaches that integrate differential privacy, homomorphic encryption, secure computation, trusted computing environments, and access control mechanisms into the unlearning process. The viability of regulatory compliance is analyzed from a sectoral viewpoint in terms of ethics, justice, and accountability, especially in high-security industries like healthcare and finance. These factors highlight the necessity of scalable, verifiable, attack-resistant, and resource-efficient MU techniques for privacy-preserving AI.