4 Topic Commentaries

Image conditioning may lower AI detectability

• 

  Hany Farid, PhD

  Digital media forensics; computer vision

  •

  University of California, Berkeley

  If I’m trying to fool people, what do I do? I generate my fake video, then I reduce the resolution so you can still see it, but you can make out all the little details. And then I add compression that further obfuscates any possible artefacts,” Farid says.

  Source
• 

  Siwei Lyu, PhD

  Digital media forensics; computer science

  •

  University at Buffalo, The State University of New York

  Watermark-based methods can be highly effective when the watermark remains intact, but any modification ​that removes or weakens the embedded signal — such as cropping, resizing, heavy compression, or ​editing — may reduce ⁠their effectiveness, depending on how the watermark is designed,” Lyu said.

  Source
• 

  Apurva Gandhi, Master’s student

  Electrical and computer engineering; adversarial machine learning

  •

  University of Southern California Viterbi School of Engineering

  If a deepfake is a virus and a deepfake detector is a vaccine, then you can think of adversarial perturbations as a mutation. Just like one tiny mutation of a virus might render a vaccine useless, tiny perturbations of an image can do the same to state-of-the-art deepfake detectors.

  Source
• 

  Shehzeen Hussain, PhD student (at time of quote)

  Electrical and computer engineering; adversarial machine learning

  •

  University of California San Diego

  Our work shows that attacks on deepfake detectors could be a real-world threat.

  Source