Privacy protection large model fine tuning method and system based on differential privacy

Abstract

The invention provides a differential privacy-based privacy protection large model fine tuning method and system. The method comprises the steps of obtaining user sensitive data to construct a fine tuning data set and a verification set; performing low-rank decomposition on the target parameter matrix, inputting the fine-tuning data set into the large model, calculating the gradient of the fine-tuning data set on the low-rank parameter matrix, and performing sparse processing; cutting the sparse gradient according to a gradient cutting threshold, and then adding differential privacy noise; correcting the variance estimated by the Adam optimizer on the basis of the gradient with the added noise and the differential privacy parameter, calculating the parameter step length according to the corrected variance, and updating the target parameter; and dynamically adjusting a gradient clipping threshold by using the optimization condition before and after updating the verification set evaluation model. According to the method, the privacy protection intensity can be ensured, and meanwhile, the large model obtained by fine tuning has relatively high effectiveness.

Images