Ant Group’s AIGC Security Detection System Earns Top Rating in China ICT Academy’s Multimodal Evaluation

China Information and Communication Research Institute recently released the first batch of evaluation results for AIGC multimodal (image/video/audio) detection services, and Ant Group’s AIGC security detection system passed all test items, receiving the highest "Excellent" rating.

With the rapid rise of generative AI tools such as Stable Diffusion, Sora, and ChatGPT, the risk of misuse—including misinformation and fraud—has grown, prompting the need for robust detection solutions.

The assessment followed the ICT Academy’s "AIGC Image/AIGC Video/Deep‑Synthesized Audio Detection Service System Requirements and Evaluation Method" standard, covering functional features, performance efficiency, service stability, and privacy protection, and evaluated more than sixty core technical indicators.

Results show the Ant AIGC security detection system can handle diverse image styles, resolutions, lighting, as well as video durations, aspect ratios, complexities, and audio scenarios, achieving a 0.99 F1 score and the top service rating in all three modalities.

The system consists of two main modules: a multimodal generation module and a multimodal understanding module, providing detection capabilities for image, video, and audio content.

According to Zhu Huijia, content algorithm director at Ant Tianjian Security Lab, the generation module covers nearly a hundred generation algorithms across modalities, enabling large‑scale synthetic data creation for training and evaluation, while the understanding module analyzes temporal, spatial, and semantic inconsistencies to verify authenticity.

Currently, Ant Group’s AIGC security solution is deployed across internal short‑video, live‑streaming, and community platforms, handling over ten million daily calls and effectively detecting deepfakes, AI‑generated rumors, and fraudulent media to protect users and maintain a healthy online environment.

To promote AIGC governance, Ant Group has open‑sourced the industry’s first million‑scale AIGC video detection dataset "GenVideo" (over a thousand downloads), announced the upcoming release of the million‑scale image detection dataset "WildFake" (accepted by AAAI 2025), and plans to release a million‑scale audio detection dataset "SpeechFake".

[1] Haoxing Chen et al., "DeMamba: AI‑Generated Video Detection on Million‑Scale GenVideo Benchmark," arXiv preprint 2405.19707. [2] Yan Hong et al., "WildFake: A Large‑scale Challenging Dataset for AI‑Generated Images Detection," AAAI 2025.