Dynamic Semantic-Constrained Adversarial Training and Defensing: A Reinforcement Learning Framework for LLM-Generated Fake Information Detection
Year: 2025
Type: conference-paper
Abstract: With the development of Large Language Models (LLMs) such as DeepSeek and GPT-4, the cost of generating false information has decreased dramatically, making detection more difficult to recognize. Traditional approaches, such as GPT-4o and chain-of-thought (CoT) frameworks, having low accuracy or efficiency. To better solve these problems, this paper introduces Dynamic Semantic Structure Rule Constrained Adversarial Training (DSC-AT), which uses balanced semantic fidelity and steganography to enhance adversarial sample optimisation, achieving an 82% deception rate. Meanwhile, the Semantic-Robust Unified Architecture (SRUA) scheme integrates text classification layer, logical verification layer and uncertainty detection layer to improve accuracy. Besides, to better train the model, we provide the Content Reinforcement Learning-based Tactical Adversarial Defence (RL-TAD) model simulates the interaction between the attacker and the defender, improves the training strategy by using backpropagation descent. Experimental results show that contrast to CoT, SRUA improves the detection accuracy to 92.1 percent and reduces the false alarm rate by 65 percent, which provided an efficient and flexible solution for identifying LLM-generated misinformation.
Authors: Yuze Ding, Shibin Zhang
Institution: Chengdu University of Information Technology
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W4411010156