Logit Laundering lets a model provider hide use of marked data

Researchers published Logit Laundering, an attack for AI providers that train on copyrighted data without permission. Data use auditing checks if data was used in training. The usual method puts watermarks (traceable signals) in the data and looks for them in model outputs. The attack weakens those signals when the model answers. The abstract gives no numbers.

How it works

The provider changes logits, the raw scores that a model gives to each next word. One module cuts the statistical signals that a watermark causes. A second module sets the strength with the confidence of the model. The model stays unchanged.

The proof

The authors say tests show that it beats the best auditing methods and keeps task results. The source gives no number for the signal drop or for a loss of quality.

What is not known

The source does not say that a known provider used it. The authors say audits must become more robust.

Insight: Attacks that keep scores steady are hard to see

Log-To-Leak, an earlier attack, kept task results in 6.63 percentage points or less in 16 of 18 pairs. Logit Laundering also keeps task skill. One more earlier paper says a private key can find a change that an attacker makes to a public signal. The two attacks hide in usual results. A score check cannot find them, and audits with a private key are possibly a stronger defence. Tell your auditor which attacks it tested.