SharedRequest: Privacy-Preserving Model-Agnostic Inference for Large Language Models

SharedRequest introduces a privacy-preserving inference framework for large language models that protects user prompt privacy by mixing prompts with noisy variants at the batch level, rather than individual-prompt level. The model-agnostic approach achieves 20% higher utility than differential privacy baselines while reducing query costs by up to 5x, requiring no modifications to LLM architecture.

SharedRequest addresses a fundamental tension in modern AI deployment: protecting user privacy while maintaining both utility and computational efficiency. As LLMs like ChatGPT process billions of prompts daily, the exposure of sensitive user queries to service providers presents significant privacy risks. The framework's innovation lies in reformulating privacy protection from individual prompts to batch-level operations, enabling cost amortization across semantically equivalent instructions without requiring architectural modifications to proprietary models.

This approach emerges from growing regulatory pressure and user concerns around data privacy. Previous privacy-preserving methods relied heavily on differential privacy techniques, which typically degrade model output quality or require computational overhead that makes deployment impractical. SharedRequest sidesteps these limitations through its model-agnostic design, making it compatible with any LLM regardless of architecture—a critical advantage given the proprietary nature of leading commercial models.

For users and service providers, the implications are substantial. Users gain privacy protections without sacrificing response quality, while providers can implement privacy without expensive infrastructure overhauls or performance degradation. The 5x cost reduction per query has meaningful implications for scaling privacy-preserving inference at production scale. Developers of privacy-focused AI applications gain a practical tool that bridges the gap between theoretical privacy guarantees and real-world deployment constraints.

The framework's success depends on batch consistency and semantic grouping accuracy. Future developments likely involve integration with production inference systems and evaluation against adversarial privacy attacks. This work contributes to making privacy-preserving AI not just theoretically feasible but economically viable for widespread adoption.

• →SharedRequest achieves 20% higher utility than differential privacy baselines while maintaining privacy protection without model-specific modifications.
• →Batch-level privacy reformulation reduces inference query costs by up to 5x through amortization across semantically equivalent prompts.
• →Model-agnostic design enables deployment on proprietary LLMs without requiring access to parameters or architectural changes.
• →Privacy-preserving inference becomes economically viable for production systems, addressing scalability challenges of prior approaches.
• →Framework supports semantic grouping of instructions to minimize utility loss while maximizing computational efficiency gains.