Privacy-Preserving Intelligence-based Reinforcement Learning for Large Language Model via Homomorphic Encryption

Reinforcement learning (RL) is a widely used framework for sequential decision-making, but conventional RL often struggles with sparse rewards, limited reasoning, and long-horizon dependencies. Intelligence-based reinforcement learning (IRL) introduces intelligence-oriented metrics and is mainly applied to large language models (LLMs) to enhance contextual reasoning and process both numerical and textual information. Nevertheless, deploying IRL with LLMs in sensitive domains like healthcare, finance, and defense raises severe privacy risks, including gradient inversion, model extraction, and sensitive data leakage. We propose a privacy-preserving IRL framework with the integration of LLMs based on the CKKS fully homomorphic encryption scheme, which supports encrypted computation on real-valued data, enabling training to be performed entirely in the encrypted domain. Security and efficiency analyses demonstrate that the framework achieves strong cryptographic security, meeting indistinguishability under the chosen-plaintext attack, and practical efficiency with low communication overhead, enabling secure deployment in privacy-critical environments despite increased computational cost.