Model Extraction of the SynthID Watermark Detector and Exploring Adversarial Attacks

The article examines Google DeepMind’s SynthID image watermark detector through two directions: attempting to extract a surrogate model from its observable behavior and exploring adversarial machine-learning attacks against the detector. The supplied metadata does not include the article’s methodology, query budget, datasets, detector outputs, attack success rates, or author conclusions. Therefore, the technical claims and reproducibility of the work cannot yet be independently assessed from the available information. It is best treated as a potentially useful security research write-up pending review of the full original text.