Invited Speakers

Abstract. Diffusion models are widely known for generating new images from text, but their capabilities extend much further, into restoration, enhancement, editing, and compression. These AI tools unlock impressive creative and technical possibilities, but also blur the line between real and synthetic content, posing new challenges for multimedia forensics. First, AI-based compression may appear content-preserving, yet it can subtly alter image semantics. This poses serious risks, for instance, when analyzing sensitive material such as surveillance footage. Moreover, such processing can introduce artifacts resembling those of generative models, meaning authentic media might be incorrectly flagged as synthetic. Even more disruptive, diffusion-based regeneration can erase forensic traces used for watermarking, deepfake detection, or image forgery localization. This can occur intentionally in targeted attacks or unintentionally through image processing such as super-resolution. Finally, current image forgery localization methods fail to localize edits performed by new AI-based inpainting methods because they fully regenerate the edited images. As AI-based regeneration becomes embedded in everyday media workflows, multimedia forensics must evolve to safeguard trust in visual evidence in the era of generative AI. 

Abstract. This talk will present recent works in the directions in face recognition driven by synthetic data and modern generative models. It will highlight how GANs and diffusion models enable the creation of realistic facial images and ID card portraits for evaluating biometric vulnerabilities, including morphing attacks and identity manipulation without relying on sensitive personal data. These synthetic pipelines offer scalable and controlled environments for stress-testing recognition systems. We also discuss generative editing techniques that modify biometric cues within latent spaces to produce natural, identity-safe outputs, supporting privacy-preserving applications while maintaining utility for downstream tasks. The talk connects progress in synthetic data generation with methods for both strengthening biometric security and preserving individual privacy.