SPS Webinar: Frequency Artefacts in Diffusion Models: an Achilles' Heel for Deepfakes?
Date: 6 May 2025
Time: 11:00 AM ET (New York Time)
Presenter(s): Dr. Quentin Bammey
Based on the IEEE Xplore® article:
Synthbuster: Towards Detection of Diffusion Model Generated Images, published in the IEEE Open Journal of Signal Processing, November 2023.
Download article: Original article is open access and publicly available for download.
Abstract
Diffusion models excel at generating photorealistic images, yet their outputs betray a hidden signature: spectral artefacts — distinct anormalies in the frequency domain that deviate from natural image statistics. This webinar will unpack the mathematical underpinnings behind these artefacts, revealing why and how incorrect frequencies emerge during the generation process. The presenter will discuss the extent to which frequency artefacts can be exploited as reliable markers to identify deepfakes, offering new tools for image forensics. Finally, he will explore strategies to mitigate these errors, paving the way for more accurate and robust image synthesis. Join us for an in-depth exploration at the intersection of signal processing and generative AI!
Biography
Quentin Bammey (M'21) received the B.S. degree in foundations of computer science, the M.S. degree in mathematics, computer vision and machine learning and the Ph.D. in applied mathematics all from ENS Paris-Saclay, Université Paris-Saclay, France, in 2021.
He is currently a researcher at the Image and Visual Representation Lab (IVRL) at École Polytechnique Fédérale de Lausanne (EPFL), Switzerland. Prior to that, he worked at Centre Borelli, ENS Paris-Saclay, Université Paris-Saclay and is mostly credited with the invention of Positional Learning, especially used to analyze the image sampling mosaic and detect forgery traces (Bammey, Quentin, Rafael Grompone von Gioi, and Jean-Michel Morel. "An adaptive neural network for unsupervised mosaic consistency analysis in image forensics." In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020).
Dr. Bammey has published several works on generated image detection, most notably the Synthbuster method and database (Bammey, Quentin. "Synthbuster: Towards detection of diffusion model generated images." IEEE Open Journal of Signal Processing (2023)). He is heavily invested in reproducible research and outreach on artificial intelligence and disinformation and participated to the coordination and development of the SaclAI-School BrevetAI pedagogical platform, to offer a learning-by-doing training on artificial intelligence. His research interests encompass image processing, computer vision, machine learning, and multimedia forensics. Dr. Bammey is a member of the IEEE Signal Processing Society and of the Systems, Man, and Cybernetics Society. He is an editor of the IPOL (Image Processing On Line, https://www.ipol.im/) journal and demo system for open science and reproducible research, and a cofounder and the organizer of the IPOL MLBriefs workshop.
