Northeastern University

Through dual representation space analysis, this work identifies an "intermediate laundering region" mechanism, builds the PADBen benchmark with a five-type text taxonomy and five progressive detection tasks, and evaluates 11 detectors, revealing a critical asymmetry: paraphrase attacks do not universally defeat detection—plagiarism evasion (paraphrasing LLM-generated text) remains detectable (RADAR sentence-pair AUC 0.909), while authorship obfuscation (paraphrasing human-authored text) collapses detection to near-random performance (AUC 0.526–0.748).

Through dual representation space analysis, this work identifies an "intermediate laundering region" mechanism, builds the PADBen benchmark with a five-type text taxonomy and five progressive detection tasks, and evaluates 11 detectors, revealing a critical asymmetry: paraphrase attacks do not universally defeat detection—plagiarism evasion (paraphrasing LLM-generated text) remains detectable (RADAR sentence-pair AUC 0.909), while authorship obfuscation (paraphrasing human-authored text) collapses detection to near-random performance (AUC 0.526–0.748).