Don't Break the Boundary: Continual Unlearning for OOD Detection Based on Free Energy Repulsion

AI systems need to both detect unusual inputs AND forget data on demand: The paper addresses a critical real-world tension—AI models deployed in open environments must identify anomalies (out-of-distribution detection) while also being able to "unlearn" specific data for privacy compliance, yet these two requirements traditionally conflict with each other.

Traditional unlearning breaks anomaly detection: When AI systems forget specific classes of data using conventional methods, they lose their ability to recognize unusual or dangerous inputs, creating a safety risk. The paper found that standard unlearning approaches cause "catastrophic loss" of the model's ability to discriminate anomalies.

The solution treats forgotten data like anomalies: Rather than simply erasing information, the proposed TFER framework reframes unlearning as transforming forgotten classes into the same category as out-of-distribution (anomalous) samples, preserving the model's safety mechanisms while still achieving privacy goals.

Efficient unlearning without full retraining: The approach uses parameter-efficient fine-tuning rather than retraining entire models from scratch, making privacy compliance and model corrections more practical and cost-effective for real-world AI deployment.

Balances competing demands through geometric precision: By using a "push-pull" mechanism inspired by physics principles, the framework keeps legitimate data in a stable region while actively pushing forgotten data to anomaly regions, maintaining both privacy protection and safety detection capabilities simultaneously.