Media Analysis, Verification and Retrieval (MeVer) Group

Our paper “The Synthetic Media Shift: Tracking the Rise, Virality, and Detectability of AI-Generated Multimodal Misinformation” was accepted at the 3rd Workshop on New Trends in AI-Generated Media and Security (AIMS) @ CVPR 2026! In this work, we introduce CONVEX, a large-scale dataset of multimodal misinformation built from X’s Community Notes, and analyze how AI-generated images and videos evolve in terms of volume, virality, engagement, consensus dynamics, and detectability. We also study how well current specialized AI-image detectors and multimodal LLMs can identify real-world AI-generated content, motivated by the growing use of AI tools within the Community Notes ecosystem. 📌 Main takeaways: • AI-generated content is becoming increasingly prevalent. • It is disproportionately overrepresented among highly viral posts. • Detecting real-world AI-generated images remains challenging: current models achieve only moderate performance overall, with multimodal LLMs performing somewhat better than specialized detectors, while both decline over time as generative models continue to evolve. This work was co-authored with Stefanos Papadopoulos and Symeon Papadopoulos, and carried out at the Media Analysis, Verification and Retrieval (MeVer) Group, Information Technologies Institute (ITI). 📄 Full paper: https://lnkd.in/d8tp8x49 🔗 Code and dataset: https://lnkd.in/dkYAa_cb

We're building an evaluation dataset for deepfake detection - because what works in the lab often breaks in the real world. Many detection tools report impressive accuracy on controlled datasets. But through our work with the WITNESS Deepfake Rapid Response Force, we’ve seen firsthand how these systems struggle with in-the-wild content: compressed videos, low resolution, noisy environments, overlapping speech, and the messy realities of social media. And yet, detection matters. When it works, it can be a powerful tool for harm reduction- helping journalists, human rights defenders, and investigators make sense of manipulated content. In our TRIED Benchmark (https://lnkd.in/e7jdDcpR), we explore this gap: showing how performance drops under real-world conditions, why more rigorous and representative evaluation is needed, and what it will take to make detection tools truly usable in practice. That’s why we partnered with Microsoft AI for Good Lab and Northwestern University to build a new evaluation dataset designed for reality, not just the lab. It brings together: • A wide range of generative AI models and manipulation techniques • Diverse, real-world (“in-the-wild”) media curated by WITNESS • Challenging scenarios that reveal where current systems fail The goal isn’t just to benchmark performance; it’s to stress-test detection systems in the environments where they actually need to work. And the need for that has never been more urgent. Our collaborative research, published in IEEE, goes deeper into the dataset design and its implications for the field: https://lnkd.in/eSR9_3XD We hope this work contributes to a more honest understanding of where detection stands today, and what it will take to make it effective in the real world. 🔎 Explore the dataset: https://lnkd.in/eKDwvhcq Grateful for our collaborators at Microsoft and Northwestern: Kevin White Juan M. Lavista Ferres Pengce Wang V.S. Subrahmanian Marco Postiglione Isabel A. Gortner and special thanks to Thomas Roca for also building such a clear, accessible dataset browser, as well as my colleagues Zuzanna W. Mahsa Alimardani Sam Gregory If you’d like to contribute to the dataset, please reach out!

Our paper "DiffusionPrint: Learning Generative Fingerprints for Diffusion-Based Inpainting Localization" has been accepted at #AIMS2026 (CVPR Workshop)! Instead of looking for noise inconsistencies, DiffusionPrint learns the generative fingerprint that inpainted regions share when produced by the same model, while remaining invariant to the artifacts introduced by latent reconstruction in authentic areas, producing a forensic feature map that plugs directly into fusion-based localization frameworks. As a drop-in replacement for noise-based modalities such as Noiseprint, DiffusionPrint integrates into fusion-based localization frameworks (e.g. TruFor, MMFusion) and consistently improves forgery localization, with gains of up to +28%. This work was co-authored with Symeon Papadopoulos 🔗 Research conducted at the Media Analysis, Verification and Retrieval (MeVer) Group, Information Technologies Institute (ITI) 🔗 Code: https://lnkd.in/dQSdujQq 📄 Paper: https://lnkd.in/d_R-pP8S

Our paper “TGIF2: Extended Text-Guided Inpainting Forgery Dataset & Benchmark” was accepted at the Journal on Information Security! 🖼️ The TGIF2 #dataset contains 271k images that were locally edited using AI. Compared to our previous TGIF dataset, we add images inpainted with the newer FLUX.1, as well as with random non-semantic masks. We benchmarked forensic methods and provide new insights. 📉 Performance drops for newer generative models ❌ Current localization methods fail on fully regenerated images. 🔧 Fine-tuning enables localization in fully regenerated images, but can introduce semantic bias. Adding random, non-semantic bias helps to remove this bias. ✨ AI-based image enhancements such as super-resolution can remove forensic traces and reduce detection performance. 🇧🇪 🇬🇷 This work was done in a collaboration between our group IDLab (UGent - imec) and Dimitris Karageorgiou, Paschalis Giakoumoglou & Symeon Papadopoulos of the Media Analysis, Verification and Retrieval (MeVer) Group at Centre for Research & Technology Hellas (CERTH). More details: 🔗 Blog post: https://lnkd.in/gDmDQ7Wz 📄 Paper: https://lnkd.in/gKUmthQi 💻 Dataset: https://lnkd.in/g7WEG6bQ

Excited to announce the 6th edition of the BIAS Workshop, co-located with ECMLPKDD 2026 in Naples. Bias and fairness in AI systems are central challenges in modern AI and are more relevant than ever, as AI moves from "controlled settings" into complex, real-world environments driven by generative AI, increasingly agentic systems, and widespread adoption. The question is no longer just "Is the model fair?" but "Does the system behave fairly?" #BIAS26 focuses on both the evolving technical nature of bias in modern AI systems, its its broader legal, ethical, and societal dimensions. We invite submissions from researchers and practitioners working on bias and fairness in AI. Submission deadline: 05 June 2026 More details and submission link coming soon - stay tuned! Organizers: MaryBeth Defrance - Ghent University Kathrin Lammers - Bielefeld University Marco Favier - University of Antwerp Symeon Papadopoulos - Centre for Research & Technology Hellas (CERTH) Eirini Ntoutsi - Bundeswehr University Munich Research Institute CODE Toon Calders - University of Antwerp #BIAS26 #AIBias #AIFairness #ECMLPKDD #ResponsibleAI #TrustworthyAI

New piece out in Tech Policy Press on how AI content detection is being weaponized in the Iran war, where claims of “fakeness” are being strategically used to discredit real documentation. "The result is a dangerous feedback loop: synthetic media undermines trust in real evidence, and fabricated forensic analysis further erodes confidence in verification itself. In this environment, the tools designed to detect manipulation become instruments of manipulation—deployed to deepen doubt and confusion about real events and even real casualties.” I feel incredibly privileged to have collaborated with the exceptional Mahsa Alimardani on this and grateful to Justin Hendrix for the care and thoughtful edits. Read more: https://lnkd.in/e4-Y9-fz

📚 What do Madame Bovary and ImageNet have in common? They are both texts! (In the semiotic sense of a coherent ensemble of signs) 🚀 In our latest article, we explore the theoretical implications on (framing) bias discovery of this semiotic characterization of image datasets 🔍 In particular, we found that framing bias is a property of image datasets that derives from their textual nature, rather than a statistical accident due to improper data collection 💡 Moreover, a consequence of our work is that the knowledge about large-scale image data is always partial, mediated and hypothetical. To overcome these issues, we propose an elastic epistemological paradigm inspired by the humanities 📖 You can find the full-text of the paper at the following link https://rdcu.be/e7RGg 📢 A shout-out to my co-authors Symeon Papadopoulos, Eirini Ntoutsi, and Yiannis Kompatsiaris

🚨 Call for Participation – Synthetic Image Detection Challenge (MediaEval 2026) We are excited to announce the second edition of the Synthetic Image Detection Challenge, organized by Media Analysis, Verification and Retrieval (MeVer) Group and AI-CODE project as part of MediaEval 2026. Following the strong participation and insights from last year’s edition, this challenge focuses on detecting AI-generated visual content and localizing manipulated regions in images collected from real-world, in-the-wild sources. 🔎 The challenge includes two tasks: Task A – Real vs. Synthetic Image Classification Detect whether an image is authentic or AI-generated. Task B – Manipulated Region Localization Identify the regions that have been modified in AI-manipulated images. ✨ What’s new in 2026 • Task A remains unchanged, enabling continuity and comparability with last year’s results • Task B introduces a new test set, reflecting more recent generative models and manipulation patterns • Stronger emphasis on analysis and insight: beyond rankings, we encourage participants to explore how evolving generative workflows and real-world transformations influence synthetic traces We particularly welcome contributions including analysis, ablation studies, and qualitative or quantitative insights that advance understanding of this rapidly evolving domain. 👩💻👨💻 We invite researchers and practitioners in computer vision, multimedia forensics, and machine learning to participate. 🔗 Learn more about the challenge: https://lnkd.in/dFsymHpJ 📝 Register for MediaEval 2026: https://lnkd.in/d6GYZ8KW 💻 Dataset & submission details (GitHub): https://lnkd.in/dFxgBf3A We look forward to seeing your contributions at MediaEval 2026! #MediaEval #SyntheticMedia #AIDetection #MediaForensics #ComputerVision #MachineLearning #MultimediaAnalysis

The Multimedia Against Disinformation Workshop, in its 5th Edition, explores AI-driven approaches for detecting, analyzing, and mitigating disinformation across text, audio, images, and video. With the rise of social networks and AI-mediated communication, disinformation campaigns, from deepfakes to LLM-generated conten, pose growing challenges to journalism, fact-checking, and public discourse. We invite contributions on topics such as: - Multimedia disinformation detection - Multimodal verification - Synthetic and manipulated media detection - Social media analysis - Robust, fair, and explainable AI tools - Large Language and Multimodal Models - Dataset development and governance - Benchmarking and evaluation frameworks - Emerging AI-driven threats …and much more! The workshop is organized in conjunction with the ACM International Conference on Multimedia Retrieval (ICMR’26), Amsterdam, The Netherlands, June 15-19, 2026. Learn more and submit your contribution: https://lnkd.in/eD6DEHr8 #Disinformation #AI #Multimedia #DeepLearning #MultimodalAI #LLM #MediaVerification #FactChecking #ICMR26 #Amsterdam #SyntheticMedia #Deepfakes #TrustworthyAI #Research 

Disinformation is no longer confined to text. It is multimodal, AI-generated, and increasingly difficult to detect. The 5th ACM International Workshop on Multimedia AI against Disinformation (MAD’26) will take place on June 15, 2026, in Amsterdam 🇳🇱, as part of ICMR 2026 (June 16–19). MAD focuses on research at the intersection of multimedia analysis and AI methods for detecting, understanding, and mitigating disinformation. This includes synthetic and manipulated media, multimodal verification, large language and multimodal models, robustness and fairness in detection systems, dataset governance, benchmarking, and emerging threats linked to the widespread use of generative AI. 🔎 Topics of interest include: - Detection of manipulated or synthetic multimedia content - Multimodal fusion and verification approaches - Multimedia forensics - Disinformation dynamics and societal impact - Robustness against adversarial and real-world conditions - Explainability for journalists and non-expert users - Large Language and Large Multimodal Models for verification - Datasets, benchmarking frameworks, and open resources 📅 Paper submission deadline: March 25, 2026 📩 Notification: April 19, 2026 📝 Camera-ready: April 25, 2026 📍 Workshop date: June 15, 2026 🌐 Website: https://lnkd.in/d5hJmppU 📨 Submission portal: https://lnkd.in/ddXuMd-s If you are working on AI methods for multimedia verification, synthetic media detection, or broader aspects of AI-driven disinformation analysis, we would be glad to receive your submission. The workshop is organized by Cristian Stanciu Roberto Caldelli Milica Gerhardt Bogdan Ionescu Giorgos Kordopatis-Zilos Symeon Papadopoulos Adrian Popescu and Dr. Vera Schmitt with the support of #AI-CODE #AI4Debunk #DeteRel and #News-Polygraph projects.