Recently, in the CVPR 2026 NTIRE Image Detection Challenge, Ant Group won the championship in both the "Robustness Sample Testing in Complex Real-World Scenarios" track and the "Face Enhancement Anomaly Detection" track, providing important support for further enhancing risk identification capabilities in scenarios such as payments, content security review, and financial identity authentication in the AI era.
Currently, the risks of deepfakes and misuse of AIGC are increasing. These fake contents are difficult to distinguish with the naked eye, and existing detection models have seen a drastic drop in accuracy when facing real-world scenarios and rapid iterations of multimodal large models. This CVPR challenge directly addresses this pain point, requiring models to maintain high accuracy and strong robustness under the dual extreme tests of "unknown generation architecture" and "complex degradation interference."
Ant Group originated from payment scenarios and has accumulated security technologies over the past 20 years that represent international leading levels. This advantage is now extended to the field of AI security. Ant Group proposed a detection framework based on the DINOv3 visual foundation model, achieving a capability leap from laboratory to real-world scenarios for AIGC detection.
In the "Robustness Sample Testing in Complex Real-World Scenarios" competition, the Ant AI Security Lab team built a complex training corpus containing millions of high-quality samples, covering open-source datasets such as WildFake, Z-Image, Seedream, Nano-banana-pro, and cutting-edge models. The underlying architecture uses a dual-stream parallel integration structure, like equipping the detection model with two complementary eyes, capturing local details and overall features of images respectively. The team simulated the full-chain degradation effects of images from single noise points to multiple distortions, deeply reproducing image distortion characteristics in real scenarios such as social platform dissemination and secondary rephotography, significantly improving the model's detection capabilities in real scenarios.
Additionally, the team proposed a two-stage detection paradigm called "Locate-Then-Examine," which first locates suspicious areas and then conducts detailed reviews, and built the FakeXplained dataset, which provides local area text explanations. When dealing with suspicious images, this method can not only accurately determine whether they are AI-generated but also locate areas with forgery flaws or violations of physical common sense on the image, while simultaneously generating detailed explanations. This approach breaks through the limitations of traditional "black-box" detection, making model decisions traceable. To facilitate technical professionals to jointly address Deepfake challenges, the team also opened source the most comprehensive AIGC image and video detection resource repository in the field via GitHub.
In the "Face Enhancement Anomaly Detection" competition, the Ant International team won by accurately locating abnormal areas in face images. This technology can precisely identify and locate abnormal areas in face images, mainly applied in scenarios such as financial transaction identity verification and account opening material review, providing important technical guarantees for preventing Deepfake and AIGC attacks. In cross-border payment and financial services, Ant International has deeply applied AIGC identification technology in EKYC, documents, and materials anti-counterfeiting, ensuring the ability to detect various types of generated content.
CVPR is an international conference on computer vision and pattern recognition hosted by IEEE, and together with ICCV and ECCV, it is known as one of the three top-tier conferences in the field of computer vision. This challenge attracted more than 500 teams from around the world.
