Saradindu Sengupta - How can a learnt ML model unlearn something: Framework for "Machine Unlearning"

Machine unlearning refers to making ML models forget specific data, features, or classes while maintaining performance

Three key pillars of machine unlearning:

• Completeness: How thoroughly data is removed/forgotten
• Timeliness: Efficiency of the unlearning process
• Verifiability: Confirming data is actually forgotten

Main types of unlearning requests:

• Individual sample removal
• Feature-based removal
• Class-wise removal
• Full dataset removal

Key challenges:

• Cost and complexity of retraining models
• Maintaining model accuracy after unlearning
• Verifying successful removal of data
• Security and privacy concerns
• Stochastic nature of ML making verification difficult

Evaluation metrics for unlearning:

• Layer-wise distance between weights
• Membership inference attacks
• Feature injection tests
• Unlearn time measurement
• Model accuracy comparison

Regulatory drivers:

• GDPR “right to be forgotten”
• FTC rulings
• Privacy regulations
• Data protection requirements

Business considerations:

• Cost of retraining vs unlearning
• Production system efficiency
• Maintaining model usability
• Meeting compliance requirements
• Balancing privacy and performance

Historical development:

• Started with kernel/SVM models in early 2000s
• Traditional ML unlearning introduced ~2015
• Recent focus on deep learning models since 2020