A hidden marker intentionally embedded in training data to detect unauthorized model training on that dataset.