Scholarlink

<p>The proliferation of advanced Large Language Models (LLMs) such as GPT-4o, Claude-3.5, Llama-3.1, and Gemini-1.5 has rendered AI-generated text increasingly indistinguishable from human-authored content, posing profound risks to academic integrity, journalistic authenticity , and information ecosystems. This paper introduces Ensemble-GAN-BERT, an advanced detection framework that synergistically combines a Generative Adversarial Network (GAN) augmented BERT architecture with a multi-model ensemble (BERT, RoBERTa, DeBERTa-v3), hand-engineered statistical linguistic features (e.g., entropy, burstiness, POS distributions), and optional watermark probing to achieve unprecedented robustness against adversarial attacks like paraphrasing. To address resource constraints, we propose a distilled variant achieving comparable performance on edge devices. Extensive evaluations on seven diverse benchmarks RAID, M4, HC3, TuringBench, and others demonstrate our framework attains state-of-the-art performance with an average accuracy of 95.8%, 0.96 F1-score, and 0.97 ROC-AUC . Notably, under strong paraphrasing attacks (e.g., T5-XXL, DIPPER), our method sustains 88–92% accuracy, outperforming baselines like DetectGPT (dropping to 4–70%) and GhostBuster (70–80%) by 15–25% margins. We further evaluate industrial tools (Turnitin, Grammarly) and complex attacks (cross-modal, long-text redundancy), revealing practical generalization gaps addressed via targeted enhancements.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Detecting the undetectable: an enhanced GAN-BERT framework with ensemble learning, statistical features, and watermarking for robust AI-generated text detection

  • 期刊论文

  • 发表日期：

    2026年6月28日

查看全文
  • Shivam Shrikant Zample,
  • Ashu Kaushik

摘要

The proliferation of advanced Large Language Models (LLMs) such as GPT-4o, Claude-3.5, Llama-3.1, and Gemini-1.5 has rendered AI-generated text increasingly indistinguishable from human-authored content, posing profound risks to academic integrity, journalistic authenticity , and information ecosystems. This paper introduces Ensemble-GAN-BERT, an advanced detection framework that synergistically combines a Generative Adversarial Network (GAN) augmented BERT architecture with a multi-model ensemble (BERT, RoBERTa, DeBERTa-v3), hand-engineered statistical linguistic features (e.g., entropy, burstiness, POS distributions), and optional watermark probing to achieve unprecedented robustness against adversarial attacks like paraphrasing. To address resource constraints, we propose a distilled variant achieving comparable performance on edge devices. Extensive evaluations on seven diverse benchmarks RAID, M4, HC3, TuringBench, and others demonstrate our framework attains state-of-the-art performance with an average accuracy of 95.8%, 0.96 F1-score, and 0.97 ROC-AUC . Notably, under strong paraphrasing attacks (e.g., T5-XXL, DIPPER), our method sustains 88–92% accuracy, outperforming baselines like DetectGPT (dropping to 4–70%) and GhostBuster (70–80%) by 15–25% margins. We further evaluate industrial tools (Turnitin, Grammarly) and complex attacks (cross-modal, long-text redundancy), revealing practical generalization gaps addressed via targeted enhancements.