Amped’s New Deepfake Detection Algorithm Published in Journal of Imaging

Dear Amped friends, we are glad to share some good news with you all. After months of complex research and testing work, we developed an effective deepfake detection algorithm. The technique can detect GAN-generated images, that are, synthesized faces “invented” from scratch by a suitably trained neural network (check this article for an introduction to deepfakes: Dealing with Deepfakes). Currently, it has been tested on images only.

GAN deepfakes are a real threat since they are often used to create fake IDs or social media profiles. What is most worrying, modern GAN deepfakes proved to be visually indistinguishable from real people’s faces, as witnessed by a recent research work by prof. Giulia Boato et al. (check our dedicated interview with one of the researchers: “More Real Than Real: a Chat About Deepfakes with Dr. Cecilia Pasquini (University of Trento“). If you want to see a few examples of impressive GAN-synthesized images, just visit https://thispersondoesnotexist.com/ and refresh the page a few times.

Amped researchers have worked hard to create a reliable GAN deepfake detection algorithm, which is, in turn, based on a suitably trained deep neural network. Indeed, detection of GAN-generated deepfakes is hardly achievable with standard image forensic algorithms since the forging network tends to emulate all the visual and statistical properties of natural images.

To put our system to the test, we took part in an international challenge about face deepfake detection.

We were provided with unlabeled testing data, we submitted our algorithm’s results, and the organizers independently evaluated the performance: our method obtained an overall accuracy of 90.05%, while the median accuracy of participants was roughly 72%.

As you know, at Amped Software, we believe that any forensic result should be reproducible. We were then happy to publish the details of how our algorithm works in this open-access scientific article, where the mentioned challenge and some of the participants’ methods are presented (our method is in Section 4.1).

Considering the encouraging detection performance, we included this new algorithm in the latest release of Amped Authenticate, 26594: you’ll find the Face GAN Deepfake filter under the Local Analysis category. Check out the related update blog post to see how the filter works!

We acknowledge that our detection system, based on Artificial Intelligence, specifically on a deep neural network, is less explainable than the other algorithms you find in our products. However, as we recently clarified in this blog, our position is that image analysis can make use of artificial intelligence, at least as a way to focus the attention of the analyst on suspicious data. Moreover, our algorithm provides both a binary decision (“Real” vs “Fake”) and a soft-valued score (a number ranging from 0 to 1.0), which tells how much the network is confident about the result. While being far from complete explainability, this is already something. Our researchers are already working on improving the explainability of our classifier’s decision, so stay tuned!