Opening remarks and welcome

Competition Introductions & Results (Track 1/2/3 Overview)

Invited Talk: Generalizable Face Forgery Detection: From Uni-modal to Multi-modal

The rapid advancement of generative AI has significantly increased the prevalence of face forgery, posing serious security and trust challenges. Face forgery detection has thus become a critical research area. In this talk, I will present recent advances and preliminary investigations in face forgery detection, with an emphasis on generalizability and multi-modal learning.

Invited Talk: From Uncertainty to Risk-Controlled Decisions: Split Conformal Prediction for Reliable AI

Modern AI systems increasingly operate in settings where high average accuracy alone is insufficient: a single hallucinated answer, missed lesion, or unsafe action may carry substantial consequences. This talk explores how split conformal prediction and conformal risk control can transform heuristic uncertainty estimates into actionable decisions with statistical guarantees. We begin with open-ended language generation, where conformal calibration enables prediction sets with coverage guarantees despite unbounded output spaces. We then examine how uncertainty distribution shift challenges standard conformal assumptions and motivate selective retention of test samples. Moving beyond set-valued prediction, we introduce risk-controlled selective answering, where AI systems learn when to answer and when to abstain. We further discuss a general framework for selection-conditioned risk control, enabling principled coordination among weak models, stronger models, and human experts through adaptive routing. Finally, we illustrate how the same calibration perspective extends to clinically meaningful objectives such as false-negative risk control in medical image segmentation. The broader message is that conformal prediction is not only a tool for constructing prediction sets, but a practical foundation for building AI systems that know when to answer, when to abstain, and when to ask for help.

Oral Paper Presentation 3

Forensic-Aware Continual Adaptation with Fisher-Guided LoRA Surgery for Image Forgery Localization

Industry Talk: Building an AI-Native Trust Foundation: Defense-in-Depth for Identity, Documents, and Agents

The rapid advancement of generative AI is reshaping digital fraud, making identity spoofing, document forgery, and agent-driven security risks increasingly scalable and difficult to detect. In this talk, I will discuss how digital trust must evolve in the AI-native era, from securing the eKYC identity entrance to verifying business documents and governing enterprise AI agents. I will introduce ZOLOZ Deeper for card AIGC risk detection, ZOLOZ ReaIDoc for multimodal document forgery detection, and the CARLI security model for agent operations. Together, these capabilities form a defense-in-depth trust foundation for resisting AI-native threats across identity, credential, and agent workflows.

🏆 Track 2 Award Winner: General AIGC Audio—Video Detection

Less is More: Modality-Decoupling for General AIGC Audio-Video Detection

Award Ceremony & Photos

Conclusion remarks

Opening remarks and welcome

Keynote speech: 从单模态到跨模态的人脸鉴伪探索

☕ Coffee Break

🏆 Track 1 Award Winner

🏆 Track 1 Award Winner

🏆 Track 2 Award Winner

🏆 Track 3 Award Winner

Award Ceremony & Photos