Intelligent Technology for Detecting Text-Based Deepfakes Using Large Language Models

Abstract

The rapid development of large language models in recent years has generated a significant problem — the increase in the volume of synthesized texts in the information environment, which poses a threat of the spread of misinformation. Accordingly, improving technologies for detecting such texts becomes a relevant ask.

This article proposes an intelligent technology for the automatic identification of texts generated by artificial intelligence, especially large language models. The research is based on the analysis of solutions from the "LLM — Detect AI Generated Text" competition on the Kaggle platform. For this purpose, a dataset was constructed that contains examples of texts from two classes: those written by humans and those generated by large language models. The dataset was compiled from data that is publicly available. An exploratory data analysis was also conducted, demonstrating the main features of the prepared dataset.

The article analyzes popular solutions for the problem of identifying texts generated by large language models within the Kaggle competition. It formalizes the general structure of the solution and justifies the main factors affecting the accuracy of identifying texts generated by artificial intelligence. An algorithm was developed to increase the accuracy of the solution through pre-processing and post-processing operations, improving the training dataset, optimizing the selection of models, and their ensemble method, among others. Experiments were conducted, demonstrating the effectiveness of the proposed intelligent technology.

This research contributes to the development of technologies to combat misinformation and highlights the importance of finding new methods to detect artificially created texts in modern information environment.