Ant Group Tops Computer Vision Conference with Breakthrough in AIGC Detection

Ant Group recently secured top honors in two tracks at the CVPR 2026 NTIRE Image Detection Challenge: "Robustness Sample Testing in Complex Real-World Scenarios" and "Face Enhancement Anomaly Detection." This achievement provides critical support for advancing risk identification in areas like payments, content security review, and financial identity authentication in the AI era.

The risks associated with deepfakes and the misuse of AIGC are growing. These synthetic contents are often indistinguishable to the naked eye, and existing detection models suffer significant accuracy declines when confronted with real-world conditions and the rapid evolution of multimodal large models. The CVPR challenge directly addresses this issue by requiring models to maintain high accuracy and robustness under extreme tests involving "unknown generation architectures" and "complex degradation interference."

With roots in payment systems, Ant Group has developed leading-edge security technologies over two decades. This expertise is now being extended to AI security. The team proposed a detection framework based on the DINOv3 visual foundation model, enabling a significant leap in AIGC detection capability from laboratory settings to real-world applications.

For the "Robustness Sample Testing" track, the Ant AI Security Lab team constructed a complex training corpus containing millions of high-quality samples. This corpus incorporated datasets such as WildFake, Z-Image, Seedream, and Nano-banana-pro, alongside cutting-edge models. The underlying architecture employs a dual-stream parallel integration structure, akin to giving the detection model two complementary "eyes" to capture both local details and global image features. The team simulated a full chain of image degradation effects, from single noise points to multiple distortions, closely replicating real-world image alterations seen in social media dissemination and secondary photography. This approach substantially improved the model's detection performance in practical scenarios.

Furthermore, the team introduced a two-stage "Locate-Then-Examine" detection paradigm. This method first identifies suspicious regions before conducting a detailed review. They also built the FakeXplained dataset, which provides localized textual explanations. When analyzing a suspect image, this technique not only determines if it is AI-generated but also pinpoints areas containing forgery flaws or physical inconsistencies, simultaneously generating detailed rationale. This breakthrough moves beyond traditional "black-box" detection, making model decisions traceable and interpretable. To foster collaboration in tackling Deepfake challenges, the team has open-sourced one of the field's most comprehensive AIGC image and video detection resource repositories on GitHub.

In the "Face Enhancement Anomaly Detection" competition, the Ant International team won by accurately locating anomalous areas within facial images. This technology is primarily applied in scenarios such as financial transaction identity verification and document review for account openings, offering crucial technical safeguards against Deepfake and AIGC-based attacks. In cross-border payments and financial services, Ant International has deeply integrated AIGC identification technology into processes like EKYC and document anti-counterfeiting, ensuring robust detection capabilities for various types of generated content.

CVPR (Conference on Computer Vision and Pattern Recognition) is an IEEE-sponsored international conference. Alongside ICCV and ECCV, it is considered one of the three premier conferences in computer vision. This year's challenge attracted over 500 teams from around the globe.