Repudiation For Good: Privacy-Preserving Prompt Deniability Through Embedding-Based Approximation

Large Language Models (LLMs) inference-time privacy focuses mainly on protecting sensitive information within the prompt. However, enabling end users to repudiate the submission of exact prompts, to avoid potential legal consequences, remains an unexplored research direction. To address this gap, we propose Repudiation for Good (R4G), a privacy-preserving approach that enables users to deny submitting exact prompts while allowing cloud-based LLM providers to receive sufficient semantic information but preventing them from proving with certainty which specific prompt was submitted. R4G is a paradigm shift in LLM inference-time privacy. Rather than encrypting exact prompts, which eventually reveal their precise content upon decryption, R4G transforms prompts into an intentionally ambiguous semantic representation through LLM embedding. This transformation creates a many-to-one mapping where multiple lexically distinct prompts collapse into a similar embedding space. The key insight is that while traditional privacy approaches aim to hide information temporarily (through encryption) or completely (through anonymization), our approach provides a probabilistic association between the prompt and its embedding, creating a deniability space, where users can legitimately deny submitting exact prompts that might carry legal, social, or professional consequences. We conducted experiments to assess the efficacy of R4G by measuring the semantic similarity between the original and approximated prompts. The results show that R4G achieved an average cosine similarity score of approximately 50%, effectively striking a balance between utility and privacy.