FALCON introduces a novel approach for machine unlearning that precisely removes sensitive information from language models while preserving model utility.

• Uses fine-grained activation manipulation rather than traditional loss combinations
• Applies contrastive orthogonal vectors to selectively target specific knowledge
• Achieves superior results in removing harmful content without degrading overall model performance
• Demonstrates significant improvements in preventing information recovery attacks

This research addresses critical security concerns by offering a more precise method to eliminate harmful or sensitive information from AI systems, reducing risks of data leaks and malicious exploitation while maintaining functionality.

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