Cross-Lingual Summarization as a Black-Box Watermark Removal Attack
Researchers tested a Cross-Lingual Summarization Attack (CLSA) by translating watermarked text, abstractively summarizing it, and optionally back-translating it. This pipeline systematically destroys token-level statistical biases in the watermark while keeping the meaning intact. The result is that detection accuracy drops to chance across multiple languages and detectors, proving CLSA is a stronger threat than simpler methods.