Tracer Newsletter #49 (14/04/20)- “Avatarify” live deepfake tool for Zoom and Skype released on Github
Welcome to Tracer- your guide to the key developments surrounding deepfakes, synthetic media, and emerging cyber-threats.
Want to receive new editions of Tracer direct to your inbox? Subscribe via email here!
“Avatarify” live deepfake tool for Zoom and Skype released on Github
Two computer vision engineers released Avartarify, an open-source tool for creating basic live facial reenactment deepfakes on Zoom and Skype.
How does Avartarify work?
The Avatarify project is heavily based on First Order Motion, a technique that animates a single photo (e.g. a picture of Elon Musk’s face), based on the motion of a driving or source video (e.g. video of a user’s face). With Avartarify, this driving video is the live footage captured via a user’s webcam on Zoom or Skype. Following installation and execution by command line, the user can simply set Zoom’s video input preferences and pick a target photo of the person to animate. The project does not require re-training or fine-tuning for each novel face found in the target photo or source video.
The growing accessibility of live deepfake creation tools
The project currently requires a powerful graphics card to operate and generates very limited outputs, particularly with the range of head and mouth movements it can replicate. However, unlike other open-source live facial reenactment techniques, it operates on mainstream video conferencing platforms and does not require re-training for each new face.
Security of video-conference software is among the top technological concerns in the current climate, and increasingly realistic real-time video impersonation will likely add to the list of vulnerabilities with remote communications in the near future. This is of particular concern when considering how cases of social engineering and fraud could be enhanced.
Other developments
1) Trump’s campaign published a misleading video splicing together two recordings of Joe Biden, making it sound like he claimed the coronavirus is a hoax. (Axios)
2) Deepmind developed a neural network for covering up voice audio glitches in Google’s video call app Duos, acting as an “auto-complete for speech”. (Google AI)
3) An artist used a first order motion model to create “living portraits” from photos of Russian world war two soldiers who died in battle. (Greetings From The Past)
4) Researchers from Google Brain and UC Berkeley published five “attack case studies” demonstrating how deepfake detectors are vulnerable to different adversarial examples. (arXiv)
5) Sophos AI used the cutting-edge language model BERT to detect malicious e-mails, such as spear-phishing attempts, where tradition spam detectors may fail. (Medium)
6) Researchers developed GANSpace, a simple technique for analyzing Generative Adversarial Networks (GANs) and creating interpretable controls for image synthesis, such as change of viewpoint, aging, lighting, and time of day. (arXiv)
7) Japanese Startup Activ8 raised one billion yen (approx $9.3m) in series C funding to further develop its Upd8 “virtual reality Youtuber” project. (The Bridge)
8) Whatsapp announced new limits to message forwarding in an attempt to limit the increased spread of coronavirus misinformation through the service. (The Guardian)
Opinions and analysis
A survey of state of the art neural rendering techniques
Researchers led by Ayush Tewari, Ohad Fried, and Justus Thies survey use cases for state of the art neural rendering techniques and explore their social implications.
How misinformation gets past social media moderators
Bellingcat’s Robert Evans outlines the techniques used by media personalities and news websites to spread coronavirus misinformation on social media without being banned.
Want to receive new editions of Tracer direct to your inbox? Subscribe via email here!
Working on something interesting in the Tracer space? Let us know at info@deeptracelabs.com
To learn more about Deeptrace’s technology and research, check out our website
