ATLAS improves model attribution when domains and versions shift
In plain words
The paper introduces ATLAS, a method for identifying which model family wrote a text, and MAABench, a set of tests for checking attribution when the writing domain or model version changes.
ATLAS learns how texts from each candidate family tend to look, then compares each text with those learned patterns while limiting the influence of a few unusually decisive words.
On MAABench’s test where both domain and version change, ATLAS assigned the correct family to 74.27% of texts, compared with 65.27% for the strongest baseline in that setting. The result applies to the five families and shifts tested.
The research
A model family can release new versions and generate text in new domains, while an attribution system may have learned from older versions and familiar writing conditions. Does attribution still work when both the domain and the version change together?
ATLAS learns how each candidate model family generates text, then compares those patterns to attribute texts under domain and version shifts.
The decisive question is whether attribution survives when both the writing domain and the version of a known model family change. MAABench tests these shifts separately and together, allowing the authors to compare performance beyond familiar training conditions.
On the joint-shift test, ATLAS correctly assigned 74.27% of texts to their source family, versus 65.27% for the strongest baseline. Both methods used the same training data and test texts, so this favors ATLAS as a complete method in that setting.
The result covers five candidate model families and the specific domain and version shifts tested in MAABench. It does not show how ATLAS handles unseen source families or text written jointly by people and models.
Analysis
Deep
- Original paper authors
- Minseok Ko, Hyunseok Lee, Woomin Song, Beomjun Kim, Jinwoo Shin
- Submitted to arXiv
- License
- CC BY 4.0