During training for image watermarking, an adversarial agent tries to remove the watermark by cropping, rotating, or adding noise to the image