RESTA (Randomized Embedding Smoothing and Token Aggregation) provides a robust defense against adversarial jailbreaking attacks that bypass AI alignment safeguards.

Key Innovations:

• Adds random noise to embedding vectors to disrupt adversarial inputs
• Performs token aggregation during generation to maintain output quality
• Significantly improves LLM resistance to harmful content extraction
• Creates more secure AI systems without sacrificing performance

This research addresses critical security vulnerabilities in modern language models, offering a practical approach to prevent malicious actors from manipulating AI systems into generating harmful content, ultimately supporting the development of more trustworthy AI technologies.

Smoothed Embeddings for Robust Language Models