Information Protection of Cross-Cultural English Translation Network Platform by Integrating Neural Network and Differential Privacy

Abstract

This paper aimed to solve the problems of user privacy leakage and insufficient platform information security in cross-cultural English translation. The goal is to design an intelligent translation network platform that takes into account both translation quality and information protection by integrating neural network and differential privacy (DP) technology. This paper collected CCMT Corpus (China Conference on Machine Translation Corpus), Europarl Corpus, OpenSubtitles, and KFTT (Kyoto Free Translation Task) corpora, and used SpaCy to process the corpus to extract key semantic information. The Transformer model uses the attention mechanism to extract sensitive semantic areas, automatically assigns weights in translation tasks, and introduces a dynamic weight adjustment mechanism to strengthen the model's attention to sensitive information. The privacy budget allocation strategy adopts a hierarchical allocation method, which is divided into high-sensitivity layer, medium-sensitivity layer, and low-sensitivity layer according to the sensitivity distribution of sentences in the corpus. Through hierarchical privacy budget allocation, Laplace noise is injected into the translation process to balance the privacy protection effect and translation quality in the translation task. Experimental results show that by combining Transformer with DP, this paper achieves a quality-privacy trade-off ratio of 143.3, 150.7, 130.0, and 147.2 in CCMT Corpus, Europarl Corpus, OpenSubtitles, and KFTT, respectively. During the 30th translation, the output standard deviation was between 0.033 and 0.042, which was very stable. Therefore, the innovative design that combines DP with Transformer improves the performance of the translation platform while protecting user privacy, providing safe and efficient technical support for cross-cultural communication.