Unlearning Attacks for Regression Learning.

Area of Science:

• Machine Learning Security
• Data Privacy

Background:

• Machine unlearning aims to efficiently remove data's influence from ML models upon request.
• Existing unlearning methods often overlook potential security vulnerabilities.
• The need for secure and robust unlearning mechanisms is critical.

Purpose of the Study:

• To propose the first unlearning attack for regression models, named UnAR (Unlearning Attack for Regression).
• To demonstrate how to deliberately manipulate the predictive behavior of regression models.
• To highlight security risks associated with machine unlearning.

Main Methods:

• UnAR misleads regression models to erase information from influential samples related to a target sample.
• Influential Sample Selection (ISS) identifies data points far from the regression plane.
• Influential Sample Unlearning (ISU) eliminates the lineage of these identified samples.

Main Results:

• UnAR successfully introduces bias into predictions for the target sample, enabling manipulation.
• Experiments on five public datasets show prediction deviations exceeding 35%.
• The attack is effective even when unlearning only 0.5% of the data.

Conclusions:

• The proposed UnAR attack poses a significant security threat to regression learning models.
• Machine unlearning processes require robust security measures to prevent malicious manipulation.
• Further research is needed to develop defenses against such unlearning attacks.