Abstract
As Large Language Models (LLMs) become ubiquitous, a significant "privacy paradox" has emerged. Users often send prompts containing sensitive information to cloud-hosted models to achieve high performance, exposing data to potential eavesdroppers. This paper introduces PFID (Privacy First Inference Delegation), a "split-and-play" framework that secures LLM interactions by sharding the model between the user and the server. By utilizing Truncated Singular Value Decomposition (SVD) to compress hidden states and a local "re-privatization" step, PFID prevents eavesdroppers from accessing sensitive data while maintaining high performance without requiring model retraining.
Introduction: The Privacy Paradox in LLM Inference
To leverage the reasoning capabilities of state-of-the-art LLMs like GPT-4, users typically transmit raw prompts to cloud servers. These prompts often contain sensitive identifiers, financial data, or passwords. This creates a vulnerability where "honest-but-curious" service providers or malicious man-in-the-middle eavesdroppers can intercept and read private information.
Existing privacy solutions, such as differential privacy or prompt tuning, often suffer from high computational costs, the need for extensive retraining, or significant degradation in model reasoning. PFID addresses these challenges by ensuring that raw user data never leaves the local device in a readable format, delegating only the heavy computational lifting to the cloud.
Methodology: Model Sharding and Re-privatization
The PFID framework employs a tripartite Model Sharding strategy, dividing the LLM into three distinct segments:
- The Head (Client-side): Processes raw input text into initial hidden states.
- The Middle (Server-side): A proprietary, high-capacity segment that performs the bulk of the computation.
- The Tail (Client-side): Reconstructs the final output tokens from processed states.
To maintain privacy during transmission, PFID uses Truncated Singular Value Decomposition (SVD). The client transmits only the top $k$ singular values—the "structural" components of the data—to the server. A critical innovation is the Re-privatization process: the client retains specific local information discarded during SVD truncation. When the server returns the processed hidden states, the client re-infuses this held-back information. This allows the local "Tail" to generate accurate text, while an eavesdropper lacking the local "key" is left with unintelligible gibberish.
Experimental Evaluation
The researchers evaluated PFID on machine translation tasks using the WMT23 benchmark, testing translations from six languages into English across three 7B-parameter models: Vicuna-7B, Zephyr-7B-beta, and Llama-2-7B-chat.
Key Results
- Performance Retention: In Chinese-to-English translation using Vicuna-7B, the original pipeline achieved a COMET score of 0.771. PFID's local decoding maintained a high score of 0.724, whereas an eavesdropper's intercepted data yielded a score of only 0.502.
- Qualitative Privacy: Case studies demonstrated that while the local client could perfectly reconstruct sensitive details (e.g., a specific financial figure like "2.68%"), an eavesdropper could only recover fragments (e.g., "..6%"), effectively losing the sensitive information.
- Error Analysis: Using GPT-4 as a judge (GEMBA-MQM metric), the study found that eavesdroppers experienced a massive spike in "critical" and "major" translation errors compared to the local PFID user.
Insights and Limitations
A key finding is that LLMs relegate specific private details to smaller singular values in the SVD process, while broader linguistic structure is held in larger values. By truncating these smaller values, PFID naturally filters sensitive data. Furthermore, the framework requires zero retraining, making it highly practical for production environments.
However, limitations remain. The framework's efficiency may be impacted by communication overhead due to multiple round-trips per token generation. Additionally, while effective for 7B-parameter models, further research is needed to determine if these sharding strategies scale to even larger architectures or non-autoregressive models.
Conclusion
PFID offers a robust, training-free solution to the LLM privacy paradox. By combining model sharding with SVD-based re-privatization, it enables secure inference delegation, ensuring that the power of cloud-based LLMs can be harnessed without compromising user data integrity.
Summary updated