IEEE SPS Webinar: Frequency Artefacts in Diffusion Models, an Achilles' Heel for Deepfakes
You may ask yourself: what do image generation supported by Artificial Intelligence (AI) and image resizing have in common? Are you interested in this, and related research questions?
If the answer is ‘yes’, this is likely to be of interest.
On 6 May 2025 at 5 pm CET, Dr. Quentin Bammey of the Image and Visual Representation Lab of the École Polytechnique Fédérale de Lausanne will give a webinar at the IEEE Signal Processing Society on the spectral artefacts left by image resampling and their use to fight disinformation.
Although diffusion models excel at generating photorealistic images, their outputs betray a hidden signature.
As images are generated in a low-resolution latent space, their decoding involves several upsampling steps that leave distinct artefacts on the generated images' spectra.
In this webinar, Quentin Bammey (formerly of École Normale Supérieure Paris-Saclay, Université Paris-Saclay, Centre Borelli - a vera.ai consortium member) will unpack the mathematical underpinnings behind these artefacts, revealing how the artefacts emerge during image generation and resampling. The extent to which they can be analysed to detect both resampled images and deepfakes will also be discussed.
If this is of interest, do join Quentin on 6 May. Participation is free of charge. However, prior registration under https://hubs.la/Q03ktwNw0 is required.
See you on Tuesday at 5 pm CET (which is 11-12 am Eastern Time)!
Author / editor: Quentin Bammey (EFPL) / Jochen Spangenberg (DW)
