Machine unlearning: linear filtration for logit-based classifiers

Area of Science:

• Machine Learning
• Data Privacy
• Computer Science

Background:

• Recent legislation grants individuals rights over personal data usage, including the "right to be forgotten".
• This presents a challenge for machine learning models trained on personal data.
• The field of machine unlearning investigates methods to remove training data from existing models.

Purpose of the Study:

• To explore machine unlearning techniques for classification models.
• To address class-wide data deletion requests.
• To propose an efficient method for sanitizing models after data retraction.

Main Methods:

• Introduction of linear filtration as a novel sanitization technique.
• Application of linear filtration to classification models, including deep neural networks.
• Experimental evaluation in an adversarial setting.

Main Results:

• Linear filtration demonstrates effectiveness in removing data influence from models.
• The proposed method is computationally efficient.
• Linear filtration shows benefits over naive deletion schemes in adversarial settings.

Conclusions:

• Linear filtration is a viable and efficient approach for machine unlearning in classification models.
• The method offers a practical solution for handling class-wide data deletion requests.
• Further research in machine unlearning is crucial for data privacy compliance.