How to Detect a Deepfake Online: Image Forensics and Analysis of Deepfake Videos
In this blog post let’s try to understand how to detect a deepfake online in a few seconds with a forensic approach. It is actually very simple by using Sensity.
With an intuitive and accessible user interface, Sensity is commodifying the technology for detecting deepfake videos and GAN-generated faces. You can easily drag & drop multiple videos and images at the same time and obtain the analysis results in a few seconds. (Format allowed: mp4, mov; png, jpeg, jfif, tiff.)
Deepfake videos detection for forensic analysis
By our records the number of fake videos online is growing exponentially since 2018, roughly doubling every six months. We have detected 85.047 fake videos as of December 2020.
To face this incredible growth, we have built a Detection Platform that uses a combination of deep learning and automated threat intelligence, able to monitor over 500 sources where the likelihood of finding malicious deepfakes is high. You can work with Sensity’s Platform by uploading your files to get a real-time analysis, or by pasting a video URL to verify its presence in our Intelligence records. Check the differences between the two approaches.
Detection by file submission:
Detection by query of our Intelligence records:
Detecting GAN generated images for forensics analysis
GAN-generated faces are inundating the web and are used with increasing frequency on social media and dating apps. With the quality of GAN-generated faces able to deceive the unprepared average Internet user, a growing number of malicious actors are leveraging their potential for scams on dating websites, promoting counterfeit products, and hiding their identity.
In the following example, a Reuters press agency’s reporter confirmed that Twitter users trying to delegitimize Russian protests pro Navalny were indeed using GAN-generated images. Sensity detector labeled these images as GAN generated with 99% confidence.
With Sensity, you can detect this kind of image and intercept scammers with confidence in the range of 95-99.9%. Additionally, investigators can obtain information on attribution, on the deep learning model used to generate the synthetic picture, e.g. Stylegan2.