Covert Prompt Transmission for Secure Large Language Model Services
Ruichen Zhang, Yinqiu Liu, Shunpu Tang, Jiacheng Wang, Dusit Niyato, Geng Sun, Yonghui Li, Sumei Sun
TL;DR
This work addresses covert prompt transmission for secure LLM services over wireless networks by jointly optimizing prompt compression and transmit power under fidelity and detection constraints. It introduces the PCAE framework, which uses surprisal-guided, head-tail token selection via a local SLM and permutation-based encryption to reduce payload while preserving inference quality; and the GPPO approach, which learns transmission policies with group-wise action sampling and KL-regularized updates to minimize latency under covertness requirements. Key results show PCAE reduces preprocessing latency by orders of magnitude while maintaining fidelity across backbones (e.g., DeepSeek-32B, Qwen-32B), and GPPO achieves up to a 38.6% reduction in covert transmission latency compared with baselines, with latency improving as power and sampling diversity increase. Collectively, the work enables real-time edge deployment of secure, undetectable LLM services, offering practical benefits for privacy-preserving and stealthy cloud-edge AI applications, while providing a principled framework for integrating semantic compression, lightweight encryption, and DRL-based covert optimization. $L_T$ denotes end-to-end latency; $F_{min}$ the fidelity threshold; and $\xi^*$ the optimal Willie's detection error, which must satisfy $\xi^* \ge 1 - \epsilon$ for covertness, all of which are handled within the PCAE-GPPO design.
Abstract
This paper investigates covert prompt transmission for secure and efficient large language model (LLM) services over wireless networks. We formulate a latency minimization problem under fidelity and detectability constraints to ensure confidential and covert communication by jointly optimizing the transmit power and prompt compression ratio. To solve this problem, we first propose a prompt compression and encryption (PCAE) framework, performing surprisal-guided compression followed by lightweight permutation-based encryption. Specifically, PCAE employs a locally deployed small language model (SLM) to estimate token-level surprisal scores, selectively retaining semantically critical tokens while discarding redundant ones. This significantly reduces computational overhead and transmission duration. To further enhance covert wireless transmission, we then develop a group-based proximal policy optimization (GPPO) method that samples multiple candidate actions for each state, selecting the optimal one within each group and incorporating a Kullback-Leibler (KL) divergence penalty to improve policy stability and exploration. Simulation results show that PCAE achieves comparable LLM response fidelity to baseline methods while reducing preprocessing latency by over five orders of magnitude, enabling real-time edge deployment. We further validate PCAE effectiveness across diverse LLM backbones, including DeepSeek-32B, Qwen-32B, and their smaller variants. Moreover, GPPO reduces covert transmission latency by up to 38.6\% compared to existing reinforcement learning strategies, with further analysis showing that increased transmit power provides additional latency benefits.