Deep fake detection

Idea name: Deep fake detection
Description: Integrate small, pre-trained NN’s into the status (Desktop) app to detect deepfakes before they are sent/received
Use case: As a user, I would like the status app to detect deep fakes and warn me:

  • when I receive a deepfake
  • before I forward them to others

Target user: Media consumer, Publishers, Content providers,

Why this is important: Social Media doesn’t exactly have a reputation to improve common good. Status has a messenger, messengers are social. Let’s set ourselves apart.

In 2017, the first deep fakes hit the web - fakes starring Obama, Putin, and Zuckerberg circulate. Research is accelerating; recent breakthroughs include researchers artificially enhancing their dance moves [1]. Facebook hosts the deepfake-detection-challenge to beef up their ability to detect fake content present in their products[2].

As a user, I would like the status app to detect deep fakes and warn me:

  • when I receive a deepfake
  • before I forward them to others

For instance, a team from Ruhr University Bochum published their take on deep-fake detection at ICML 2020, which maintained deepfake artifact detection accuracy of over 90% under adversarial conditions[3][4].

[1] https://www.youtube.com/watch?v=mSaIrz8lM1U&feature=youtu.be
[2] Deepfake Detection Challenge Results: An open initiative to advance AI
[3] https://arxiv.org/pdf/2003.08685.pdf
[4] GitHub - RUB-SysSec/GANDCTAnalysis: Code for the ICML 2020 paper: Leveraging Frequency Analysis for Deep Fake Image Recognition.

This is an exciting way to reinforce Status’ mission! Could deepfake checks be done by a desktop Status app if it’s synchronized to by the mobile app, perhaps as a way to encourage users to install the desktop client? Would users get something like “Deepfake: 89% likely” if images reach a certain threshold, perhaps one that users can customize if they opt-in?

2 Likes

This topic was automatically closed 182 days after the last reply. New replies are no longer allowed.