HarmAug solves the challenge of deploying safety guard models on resource-constrained devices through innovative knowledge distillation techniques.

• Introduces a data augmentation approach specifically designed for security-focused knowledge distillation
• Creates smaller safety models (125M parameters) that achieve 96% effectiveness of larger models (7B parameters)
• Enables deployment of safety guardrails on mobile devices with minimal memory and latency costs
• Demonstrates that specialized techniques outperform standard distillation for security applications

This research is critical for security as it allows protective guardrails to be deployed broadly across devices, preventing harmful instructions and jailbreaking attempts without requiring massive computational resources.

HarmAug: Effective Data Augmentation for Knowledge Distillation of Safety Guard Models