LoTUS: Large-Scale Machine Unlearning with a Taste of Uncertainty

LoTUS is a novel machine unlearning method that removes the influence of training data from pre-trained models without requiring full retraining. The approach smooths prediction probabilities to reduce over-confidence from memorized data and introduces a new evaluation metric (RF-JSD) for real-world conditions, outperforming existing methods on large-scale datasets like ImageNet1k.

LoTUS addresses a critical challenge in machine learning: the ability to efficiently remove specific training samples from pre-trained models without discarding the entire model and retraining from scratch. This capability matters because it enables compliance with data deletion requests (as required by privacy regulations like GDPR), mitigates risks from poisoned or sensitive training data, and reduces computational waste in model maintenance.

The research emerges from growing recognition that large models retain memorized information about training samples, creating privacy and security vulnerabilities. Previous unlearning approaches either required expensive retraining or achieved incomplete data removal. LoTUS tackles this by smoothing the model's prediction probabilities toward an information-theoretic bound, effectively reducing the model's confidence in ways that stem from data memorization rather than genuine learned patterns.

For the AI and machine learning industry, this work has practical implications. Large technology companies and AI developers face increasing pressure to demonstrate data deletion capabilities. The introduction of RF-JSD as an evaluation metric enables more realistic assessment under production conditions, where retraining is infeasible. The evaluation on ImageNet1k—a genuinely large-scale dataset—signals that the method scales to real-world model sizes and complexities.

The broader significance lies in enabling responsible AI deployment. As regulatory frameworks tighten around data rights and model accountability, efficient unlearning becomes infrastructure-level capability. Organizations can now consider retention policies and data lifecycle management as engineering problems rather than theoretical exercises.

• →LoTUS eliminates training sample influence from pre-trained models without full retraining, addressing privacy and security concerns.
• →The method smooths prediction probabilities to mitigate over-confidence stemming from data memorization.
• →Introduction of RF-JSD metric enables evaluation under real-world conditions where retraining is impractical.
• →Experimental validation on ImageNet1k demonstrates scalability to large-scale datasets with practical efficiency gains.
• →Open-source release enables broader adoption and integration into production ML pipelines.