FALCON introduces fine-grained activation manipulation techniques that allow precise removal of sensitive information from large language models while preserving overall performance.

• Uses contrastive orthogonal unlearning to target specific knowledge rather than broad categories
• Achieves superior balance between information removal and model utility preservation
• Operates at the activation level for more precise control than traditional loss-based approaches
• Addresses critical security concerns by preventing harmful information retention in AI systems

This research matters for security professionals as it provides a more reliable method to ensure AI systems cannot retain or expose sensitive data, significantly reducing potential security vulnerabilities in deployed models.

FALCON: Fine-grained Activation Manipulation by Contrastive Orthogonal Unalignment for Large Language Model