Ant Security's Tianjian Content Risk Control System Receives Five‑Star Rating in 2022 Content Review Service Evaluation

On January 17, the China Academy of Information and Communications Technology (CAICT) and the Content Technology Industry Promotion Alliance released the "2022 Content Review Service System Evaluation" results. Ant Security's self‑developed Tianjian content risk control system received the highest five‑star rating in both text and image core assessment scenarios.

Currently, building a healthy online content ecosystem has become a key societal focus. The rapid emergence of internet applications lowers the barrier to content creation, boosting productivity but also introducing challenges such as false information, online harassment, privacy leaks, piracy, and vulgar or obscene content.

Since 2017, CAICT has been researching content review technologies, collaborating with the China Communications Standards Association, leading domestic internet companies, and major industry enterprises to develop the "Network Governance Capability Assessment Specification: Part 2 – Harmful Information Content Identification Service System," which evaluates intelligent recognition functions and performance for images, text, audio, and video. In 2022, CAICT updated and enriched the test datasets to reflect current technological developments, making the evaluation highly recognized and benchmarked within the industry.

Ant's Tianjian risk control system leverages billion‑scale data to build underlying algorithm models, moving beyond traditional manual or static detection methods. It achieves comprehensive multimodal perception across text, image, audio, video, and behavior, creating an intelligent and efficient risk control engine that meets real‑time, large‑scale platform demands and safeguards cyberspace security.

The text algorithms rely on accumulated domain knowledge and risk corpora, employing natural language understanding, knowledge graphs, and adversarial robustness techniques to accurately identify unconventional content such as machine‑generated text, variant characters, cartoon distortions, and obfuscated expressions. The image algorithms use reinforced layout, multimodal pre‑training frameworks, and multi‑label classification to build a universal factual risk tag recognition capability, which can be flexibly combined and adapted to various application scenarios.

Looking ahead, the Tianjian system will strengthen multimodal content understanding, develop technologies for heterogeneous data fusion, cross‑modal association, and controllable generation, and create an extensible algorithm framework that can quickly respond to new content types and emerging risks, delivering tailored risk‑control solutions for diverse industry contexts and contributing to a mutually beneficial, open, and inclusive content ecosystem.