You don't even need an adversarial attack method to fingerprint or watermark a proprietary dataset. If the dataset consists of images, you could just watermark

You don't even need an adversarial attack method to fingerprint or watermark a proprietary dataset. If the dataset consists of images, you could just watermark them, or mark them using steganography. The watermark would appear as random (non-targeted) noise, and would be mostly invisible to current classifiers. Please note that the noise employed by adversarial attacks is very different, it is highly targeted.