StealthRL: Reinforcement Learning Paraphrase Attacks for Multi-Detector Evasion of AI-Text Detectors

1. 1Train a paraphrase policy using Group Relative Policy Optimization (GRPO) with LoRA adapters
2. 2Optimize the policy against a multi-detector ensemble to evade detection
3. 3Use a large language model like Qwen3-4B as the base model for paraphrasing
4. 4Evaluate the robustness of AI-text detectors under realistic adversarial conditions

AI engineers and researchers benefit from StealthRL as it helps stress-test detector robustness, while product managers and security teams can use it to improve the security of AI-text detectors

💡 Reinforcement learning can be used to generate paraphrases that preserve semantics while evading detection by AI-text detectors

🤖 StealthRL: a reinforcement learning framework for generating paraphrases that evade AI-text detectors #AI #AdversarialAttacks