Verifying whether a model was trained on a specific dataset by detecting planted signals in model outputs.