2026-11-09 · America/Los_Angeles · 社区动态 · #19

Would you watermark your dataset if you could later check whether a model trained on it?

I’ve been looking at research on radioactive data watermarks (TRACE, STAMP, Meta’s radioactive watermarking). The idea is that you rewrite a dataset with a hidden statistical pattern tied to a secret key. If a model is later trained on it, its outputs carry the pattern, and you get a p-value for this model saw your data. Models that never saw it don’t show the pattern. As far as I can tell, there’s no easy tool for doing this with text datasets, so I’m considering building an open-source one. Before I do, I’d like to know if it would actually be useful. If you publish datasets or benchmarks…

热度 43.8 / 100；排名与评分保留该期记录。