Detectors that perform well on raw LLM output often collapse once the text is paraphrased, “humanized” by adversarial tools, or lightly edited. This thesis studies how robust current detectors are to realistic evasion strategies and whether adversarially-trained or ensemble detectors hold up better.

Tasks

• Assemble a testbed of evasion strategies: automatic paraphrasing, back-translation, style-transfer “humanizers,” and manual light-editing.
• Benchmark leading detectors (watermark-based, statistical, classifier-based) against each attack type.
• Explore mitigation: adversarial fine-tuning, ensembling signals (stylometric + statistical + classifier), or robustness-aware training.
• Report the accuracy/robustness trade-off and where each detector family breaks down.

Prerequisites

• Python
• solid grounding in NLP/ML evaluation methodology
• some familiarity with adversarial ML is a plus

Why this thesis is special

Evasion is the practical failure mode that matters most for real deployments (academic integrity, misinformation moderation); systematic robustness studies here are still sparse.