German researchers have found a way to identify deep fake images by converting them to the frequency domain using the discrete cosine transform.
The team, from the Ruhr-Universität Bochum and Casa (Cyber security in the age of large-scale adversaries), point those sceptical about deep fakes to the website www.whichfaceisreal.com – a website whose data set was used as a resource for the research.
The fakes are generated with the help of ‘generative adversarial networks (GANs), according to the university, where two algorithms work together in a loop: the first algorithm creates random images based on some initial data, and the second algorithm assesses the generated image, sending it back to the first algorithm for revision if it can be identified as a fake, and so on.
“To date, deep-fake images have been analysed using complex statistical methods,” according to the university. “The Bochum group chose a different approach by converting the images into the frequency domain using the discrete cosine transform. The generated image is thus expressed as the sum of many different cosine functions.”
Images of people transformed into the frequency domain – upper left corner represents low-frequency image areas, the lower right represents high-frequency areas.
The key finding was that natural images consist mainly of low-frequency functions (left), while images generated by GANs exhibit artefacts in the high-frequency range – for example, a grid structure (right) emerges in the frequency representation of fake images – artifacts caused by the up-sampling operations found in all current GAN architectures.
“Our experiments showed that these artefacts do not only occur in GAN generated images. They are a structural problem of all deep learning algorithms,” said system security researcher Joel Frank. “We assume that the artefacts described in our study will always tell us whether the image is a deep-fake image created by machine learning. Frequency analysis is therefore an effective way to automatically recognise computer-generated images.”
The team presented its work at the virtual International Conference on Machine Learning (ICML) yesterday. The paper, ‘Leveraging frequency analysis for deep fake image recognition‘, has also been published on-line.
“Additionally,” said the university, “the researchers make their code freely available on-line, so that other groups can reproduce their results”.