Deepfake Detectors Can Be Defeated, Computer Scientists Show for the First Time

Title is sensationalist, while the news really ain't a big deal IHMO. Adversarial attacks against Neural Networks, or any other heuristics for that matter, are widely recognized in the industry.

Frameworks/papers aiming at benchmarking and fine tuning networks against those attacks even exist (e.g. github.com/QData/TextAttack in the case of NLP).

The tools will simply be adapted. Either to a/ identify the statistical anomalies/artifacts introduced by the manipulation process, or b/ ignore the manipulation through further training or further data normalization.

What makes this headline troubling to me is it allows for a conspiracy of “they are just labeling this non deepfake video supporting my side as an adversarial deepfake so people will ignore it”

In general deepfake technology exacerbates the epistemological crisis we are in, wondering how we can know what the true story is.

This isn't news. The whole idea of adversarial algorithms is to fight detector vs. creator (or some other difference in sides). Deepfake detectors can be defeated, so we'll make better deepfake detectors. Then better deepfake creators come along to beat those.

A new arms race begins.