🌳 METR builds on the Tree-Ring watermark by Yuxin Wen and John Kirchenbauer, overcoming its inability to encode meaningful messages, while having high robustness 🔒 We also combined METR with Stable-Signature by Pierre Fernandez and Guillaume Couairon to be able to add more unique messages 2/

The main challenges that modern watermarking algorithms face are with their ability to endure attacks and encrypt many unique messages, such as user IDs. METR: Message Enhanced Tree-Ring is an approach that aims to address these challenges. 4/

METR Watermark is concentric circles in Fourier space of initial noise. We encrypt binary messages by converting 1s to circles with value "S" and 0s to circles with value "-S". To detect a watermark, we take the sign of the mean value on each circle on the predicted noise. 5/

METR++ allows encryption of virtually unlimited amount of unique messages by combining METR with Stable-Signature by Pierre Fernandez. We group users into groups of size 2^R, which is the maximum messages METR can handle, and encode each group with Stable-Signature watermark. 6/

We show METR's robustness to attacks and compare it to Tree-Ring for the detection of presence of watermark. It is hard to correctly compare these methods, because Tree-Ring cannot encrypt any meaningful messages, while METR is able to encode up to multiple thousands. 7/

METR is noticeably more robust than Stable-Signature, even though Stable Signature can only add a single message per pretrain. While METR++ inherits low robustness to attacks from Stable-Signature, it has a possibility to create unlimited number of unique messages. 8/

Image quality degradation is not that big for the ability to encrypt multiple messages. And it also can be regulated with the decrease of the radius or properly selected value of S. 9/

Summary: Selection between METR and METR++ creates a robustness-number of messages tradeoff. If a virtually unlimited number of messages are needed, one might have to use METR++ and if high robustness to attacks is needed, it is better to stick to the METR watermark. 11/

Just curious if there is any work, that shows, that data from one modality can improve performance in the other?