EU co-funded projects AI4Media and vera.ai co-organized a joint workshop on Multimedia AI against Disinformation (MAD’23). It took place in Thessaloniki on 12 June 2023. Here's a recap and outlook.
The program (available on the workshop website), was exciting and intense, attracted a diverse audience of researchers on AI and multimedia, and featured two keynote speeches from highly renowned researchers of the AI4Media project.
After a warm welcome by the workshop organizers (among them IDMT's Luca Cuccovillo and CERTH/ITI's Akis (Symeon) Papadopoulos, both involved in AI4Media and vera.ai), Prof. Roberto Caldelli started with a keynote about lessons learned and new challenges in multimedia forensics. The talk provided a summary of the evolution of manipulations in images and videos, starting from the “traditional” field of image and video forensics. It critically discussed how the newly emerging field of synthetic media detection is different compared to that of media forensics.
The first session on “AI for Audio Analysis” started with a work on synthetic speech detection through audio folding by ISPL (Politecnico di Milano), and continued with a presentation on a spoofing-aware transformer network by SECS (Oakland University). Both works proposed new innovative approaches to detect synthetic speech content generated by the latest neural network technologies.
The second session on “Improving AI Generalization” continued the workshop with work on autoencoder-based data augmentation for synthetic image detection by the AI Multimedia Lab (UPB). This was followed by a presentation on quality-based training content selection for synthetic image detection by the MeVer group (CERTH-ITI), and concluded with a presentation on using “synthetic misinformers” to combat multimodal misinformation, also delivered by members of the MeVer group (CERTH-ITI). All three works and respective presentations dealt with the problem of generalization, i.e., on training neural networks which can address fake content generated by models and processes that are not known at training time.
After a lunch break full of exchanges and discussions, Prof. Ioannis (Yiannis) Patras held a keynote on controllable image generation and manipulation. The talk provided the audience with an overview on the latest advances in AI-generated visual (image/video) content, underlining the new research trends of the next-generation of synthetic content to come – on the light motive of “to know your enemy, you must become friends with them” (kidding - Ioannis Patras is among the good guys 🙂).
The third and last session on “AI for (Dis-)Information Analysis” included two works from a social science perspective, addressing two highly debated topics: the coverage of the current conflict in Ukraine was addressed by an independent Deloitte consultant (Benjamin Shultz), whereas the debate on the Covid-19 pandemic and its impact on media was thoroughly discussed by the IDIAP Social Computing Group.
The workshop ended with a (much awaited!) open discussion on MAD challenges and opportunities: organizers, researchers presenting their works, and the audience attending the event joined together to provide feedback on the day, to discuss the current status of research and the difficulties/challenges connected to it. They also highlighted the opportunity of organizing a new edition of the MAD workshop in 2024.
We - the organisers - were very pleased with the large attendance, and are grateful that the entire event was held completely in person. Everyone had the chance to contribute, network, and make new friends. It was intensive (very much!) but worth all the effort invested.
Are you as eager to know and find out - as participants were - where the next MAD workshop will be held? Below is a little trivia for you – AI generated, of course 😉.
MAD’24 teaser / prompt: “expressive oil painting of a beach in <SUPER-SECRET-LOCATION>, using warm colors”
All contributions to the workshop are available online on the ACM website. You can find them following this link or by inserting the URL https://dl.acm.org/doi/proceedings/10.1145/3592572 into your browser.
Author: Luca Cuccovillo (Fraunhofer IDMT)
Editors: Jochen Spangenberg (DW), Akis Papadopoulos (CERTH-ITI, mever team)