LLM Unlearning: Cyber Defense for Sensitive Data

Large Language Models (LLMs) are increasingly integrated into sensitive applications across various sectors, yet their inherent inability to "forget" poses significant cybersecurity, privacy, and safety challenges. Information such as personal data, copyrighted material, or hazardous knowledge can remain embedded within the model's parameters, making LLMs vulnerable to data extraction, jailbreak attacks, and regulatory breaches. Real-world incidents underscore the urgency of this problem. Retraining massive LLMs to remove specific information is computationally impractical. Consequently, LLM unlearning has emerged as a primary cyber defense mechanism. This approach aims to selectively eliminate or suppress targeted knowledge from a trained model without a full retraining cycle, while preserving its general capabilities. The survey primarily focuses on gradient-based unlearning methods due to their compatibility and scalability. A central, unresolved question remains: do these methods genuinely erase knowledge, or do they merely prevent its expression under typical prompting conditions?