How AI Large Models Are Revolutionizing Multimodal Content Safety

The award‑winning joint project between Shanghai Jiao Tong University and Ant Group, titled “AI Large‑Model‑Driven Multimodal Content Safety Governance Key Technologies and Applications,” received the first‑class Science and Technology Progress Award at the 2025 China Image and Graphics Conference.

The team built a foundation model for full‑modal content risk cognition, focusing on challenges such as deep multimodal understanding and covert knowledge injection, thereby systematically enhancing the model’s ability to recognize complex multimodal content.

In parallel, they created highly sensitive, continuously evolving detection techniques for malicious forged content, solving the difficulty of detecting covert harmful outputs from generative AI (AIGC) systems.

A new generation intelligent content risk‑control architecture centered on “instruction following” was proposed, which significantly boosts response efficiency to emerging risks without increasing inference costs.

The solution is already deployed in Ant Group’s Tianjian Intelligent Content Risk‑Control platform, protecting massive data streams from services such as Alipay Wealth Community, live streaming, and short‑video platforms, and has markedly improved defense against sophisticated adversarial risks.

Ant Group has established nine security labs for frontier research and open collaboration, forming a comprehensive security defense layout. The China Image and Graphics Society, founded in 1990, awards the CSIG Science and Technology Prize to promote innovation and industrial development in the field.