The latest articles on DEV Community by Courtlyn Deitch (@deitch83919). https://dev.to/deitch83919 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3994720%2F04cf8998-77ca-4e00-ab27-1ec5e5993213.png https://dev.to/deitch83919 en Courtlyn Deitch Wed, 24 Jun 2026 03:24:40 +0000 https://dev.to/deitch83919/notes-on-adversarial-paraphrasing-a-paper-review-214o https://dev.to/deitch83919/notes-on-adversarial-paraphrasing-a-paper-review-214o <p>Just finished reading Saha et al. arXiv 2506.07001 on adversarial paraphrasing for AI detector evasion.</p> <p>Key claim: detector-guided paraphrasing with RoBERTa as reward reduces TPR by 87.88 percent across Binoculars, Fast-DetectGPT, Ghostbuster, RADAR, GPTZero. Universal, training-free.</p> <p>What surprised me: the approach works even on detectors that were trained with adversarial examples baked in. Suggests the discriminator signal is fundamentally narrower than the generator space.</p> <p>Open questions:</p> <ul> <li>Does this generalize to detectors using surprisal variance (DivEye 2509.18880)?</li> <li>Multi-LLM round-robin generation: would mixing 3-4 models in pipeline give even more headroom?</li> <li>Token-level homoglyph substitution (SilverSpeak) is trivially detectable via Unicode normalization, but adversarial paraphrasing leaves no such forensic signal.</li> </ul> machinelearning nlp ai security