Combining Expert Knowledge with LLMs for Generating Privacy-Computing Code

PDF view of the paper entitled Spdzcoder: Combining expert knowledge with LLMS to generate a code for a privacy holder, through XiaONING Dong and 2 other authors

PDF HTML (experimental) view

a summary:Private computing receives increasing attention, but writing the computing code for privacy still represents a challenge for developers due to the limited library functions, which requires the implementation of the job from the zero point, data requirements, contradiction of the intuitive thinking and usual practices of programmers. The automation of the computing symbol generation with privacy with large language models can simplify the development effort and reduce the barrier in front of the use of private computing frameworks. However, the current LLMS still faces challenges in translating the code for the conservation of privacy, such as Python translation into MP-SPDZ, due to the scarcity of MPDZ data required for pre-training or control. Moreover, the lack of a standard increases the complexity of the translation quality assessment. To treat restrictions, this work suggested SPDZCOMER, a bases -based framework that combines LLMS and experts knowledge to generate a symbol for a privacy holder without the need for additional training data. Specifically, SPDZCoder uses a strict procedure to collect high-quality experts to represent semantic differences between Python and MP-SPDZ, and derives the rules of translation to translate Python into MPDZ based on this knowledge. Next, SPDZCoder gradually converts the Python icon to MP-SPDZ icon using the shift rules in a three-stage pipeline. To evaluate SPDZCoder, we manually built a standard data set, SPDZEVAL, which includes six data splits, each of which represents a distinctive class of difficult tasks in the MPDZ implementation. Wide experiences show that SPDZCoder achieves great performance, and greatly exceeds the foundation lines in Pass@1 and Pass@2. Specifically, SPDZCoder achieves a comprehensive health of 85.94 % and 92.01 % in the@1 and Pass@2, respectively, while the best performance baseline achieves 63.58 % and 76.36 %, respectively.