Ant AI Wins CVPR 2026 Challenge: A Powerful Countermeasure Against Deepfake Abuse

Deep‑fake and face‑swap technologies have increasingly been abused in the entertainment industry, with incidents such as AI‑generated short dramas that impersonate actors like Yang Zi and Yi Yangqianxi without permission, prompting public outcry and an official statement from the China Broadcast Television Social Organization condemning illegal AI manipulation.

At the CVPR 2026 NTIRE Robust AI‑Generated Image Detection in the Wild Challenge, the Ant Group AI Security Lab team (MICV) achieved a ROC AUC of 0.9723 on robustness test samples, securing the championship among more than 500 participating teams.

The team also released the open‑source repository Awesome‑AIGC‑Image‑Video‑Detection , which aggregates recent incidents, cutting‑edge papers, benchmark datasets, and practical tools for the global research community.

The challenge highlighted two core difficulties: (1) insufficient cross‑domain generalization due to diverse generation architectures (diffusion, autoregressive, closed‑source platforms) and (2) real‑world degradation (compression, blur, noise) that obscures subtle forgery cues.

To address these, the team built a robust detection framework based on the DINOv3 visual foundation model. The framework includes hierarchical data construction with a multi‑source corpus of millions of samples covering open‑source benchmarks, targeted synthetic data from mainstream generators, high‑fidelity commercial API samples, and official challenge data.

The architecture employs a dual‑stream parallel design: two independent DINOv3 backbones extract multi‑scale local and global features, which are fused via a multi‑scale aggregation mechanism and fed to an MLP head. Predictions from both streams are combined by weighted averaging, preserving complementary information while avoiding early feature dilution.

Robust data augmentation simulates a stepwise degradation pipeline covering blur, noise, compression artifacts, color shift, and geometric distortion, as well as high‑fidelity compression models (HiFiC, ELIC) and real‑world social‑media artifacts. Model optimization techniques such as Focal Loss, Stochastic Weight Averaging (SWA), and Test‑Time Augmentation (TTA) further improve performance on degraded samples.

Beyond the DINOv3 framework, the team introduced the Veritas framework, which integrates a multimodal large language model (MLLM) with pattern‑aware reasoning to enhance cross‑domain inference, and the Locate‑Then‑Examine two‑stage detection paradigm that first localizes suspicious regions and then conducts detailed analysis, reducing hallucination and boosting precision.

These technologies have been deployed across Ant Group platforms (short video, Lingguang, Jingtan, etc.), serving billions of users and earning certifications such as CNAS and iBeta. The Ant AI Security Lab’s continued research has resulted in over 50 international patents, 8 high‑impact papers since 2024, and multiple world‑championship titles in AI forensics competitions.