SecPE introduces a groundbreaking fusion of private inference and prompt ensembling to create LLMs that protect user data while maintaining resilience against attacks.

• Employs Fully Homomorphic Encryption (FHE) to encrypt user prompts during inference
• Implements prompt ensembling to enhance resistance against adversarial attacks
• Achieves both privacy preservation and robustness simultaneously, addressing two critical security concerns
• Demonstrates minimal performance degradation while providing substantial security benefits

This research addresses crucial enterprise concerns about deploying LLMs, offering a pathway to secure AI systems that can protect sensitive data while maintaining reliability against malicious exploitation.

SecPE: Secure Prompt Ensembling for Private and Robust Large Language Models