Scholarlink

Removing the effect of a specific subset of training data from a fully trained machine learning model is critical in many real-world scenarios. However, retraining the entire model from scratch is often impractical and resource-intensive. Therefore, efficient methods for removing the influence of selected training data without full retraining are highly desirable. In this paper, we propose a novel algorithm that effectively neutralizes the influence of a designated data subset on an already trained model. Our approach generates targeted noise that counteracts the impact of the chosen subset while preserving the model’s performance on the remaining data. Empirical evaluations across various deep learning models (ResNet9, ResNet18, AllCNN, and MobileNetV2) and datasets (MNIST, CIFAR-10, SVHN, and CASIA-WebFace) demonstrate the effectiveness and efficiency of the proposed method. The source code is available at  https://github.com/rjdpm/anti-samples .

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

A Novel Anti-Sample Generation Technique for Effective Machine Unlearning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Rajdeep Mondal,
  • Soumitra Samanta

摘要

Removing the effect of a specific subset of training data from a fully trained machine learning model is critical in many real-world scenarios. However, retraining the entire model from scratch is often impractical and resource-intensive. Therefore, efficient methods for removing the influence of selected training data without full retraining are highly desirable. In this paper, we propose a novel algorithm that effectively neutralizes the influence of a designated data subset on an already trained model. Our approach generates targeted noise that counteracts the impact of the chosen subset while preserving the model’s performance on the remaining data. Empirical evaluations across various deep learning models (ResNet9, ResNet18, AllCNN, and MobileNetV2) and datasets (MNIST, CIFAR-10, SVHN, and CASIA-WebFace) demonstrate the effectiveness and efficiency of the proposed method. The source code is available at  https://github.com/rjdpm/anti-samples .