AI Watermarking 101: Tools and Techniques

Hugging Face · 2 years ago · 31

Watermarking techniques are being developed to mark AI-generated content such as images, text, and audio in order to authenticate their origins and prevent non-consensual manipulation. Tools like IMATAG and Truepic embed imperceptible watermarks in images during or after generation, while text watermarking works by restricting or promoting certain tokens during language model output, and AudioSeal watermarks speech at frequencies imperceptible to human hearing. The adoption of these methods faces tradeoffs between making watermarking systems open to foster innovation versus keeping them closed to prevent malicious actors from learning how to remove or circumvent the watermarks.