Methods and systems for protecting privacy of large model inference data in confidential computing environments

Abstract

This invention relates to a method and system for protecting the privacy of large model inference data in a confidential computing environment. The method includes: real-time acquisition of the raw token stream generated by model inference, recording the conditional probability and generation time, and performing semantic correlation analysis to generate a sequence of raw token tuples; calculating information entropy weights based on conditional probabilities and appending them to the corresponding token tuples to generate a complete token tuple sequence; dynamically aggregating and planning the tokens based on semantic equivalence class identifiers and information entropy weights to generate a token block partitioning scheme; re-encapsulating the tokens based on the partitioning scheme to generate a token block sequence; and performing length standardization processing on each token block to generate network transmission packets with uniform length characteristics and sending them to the user terminal. This invention solves the problem that existing trusted execution environments cannot prevent network side-channel attacks by eliminating the statistical correlation of data packet timing and size through semantically preserved traffic shaping, while ensuring the semantic coherence of the output.

Images