Booster is a new defense mechanism that protects large language models from malicious fine-tuning attacks by identifying and attenuating harmful weight perturbations.

• Identifies harmful fine-tuning by targeting specific weight perturbations that break model alignment
• Achieves 90%+ success rate in defending against alignment-breaking attacks
• Maintains original model performance while blocking harmful behavior
• Offers stronger protection than existing defensive strategies

This research addresses a critical security concern for AI providers offering fine-tuning services, helping maintain trustworthy and safe AI systems in deployment.

Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbation