AI-Powered Real-Time Content Risk Monitoring for Alibaba Live Streaming during Double 11

Live streaming has become a major interactive format and a highlight of Alibaba's Double 11 event, but its content risk monitoring poses significant technical challenges, including the lack of mature solutions, uncontrolled streamer behavior, massive concurrent streams, and the need for real‑time algorithmic responses.

Alibaba's security division addressed these challenges by extensively applying artificial intelligence and deep‑learning techniques alongside an optimized high‑performance multimedia computing cluster, enabling the system to process 5,400 live video streams and 250,000 fan‑interaction games during peak periods while issuing warnings or blocking violations.

The solution focuses on two core technologies: real‑time content filtering and multimedia processing cluster optimization. For real‑time filtering, analysis of historical violation records revealed that pornographic and politically sensitive facial content dominate the risks. Consequently, two algorithmic services—video porn detection and sensitive‑face detection—were deployed, achieving roughly 99% automatic review with only about 1% of videos escalated to human auditors.

The intelligent porn detection engine assigns a score from 0 to 100 to each image or video frame, where scores ≥99 indicate definite pornographic content, 50‑99 require human review, and <50 are considered safe. Built on a multi‑layer visual perception network using an improved Inception architecture and model cascading, the engine reaches 99.6% accuracy, processes over 6 million labeled images, and reduces manual review to roughly 0.1% of cases.

Sensitive‑face detection handles 1‑to‑N matching of faces against a database of ~20,000 well‑known individuals, using a deep‑learning model with five convolutional layers and two fully‑connected layers, trained with multi‑task loss functions to capture age, gender, and other attributes. The system can return the most similar face, name, and similarity score within 10 ms for million‑scale datasets, achieving a recall rate of 93.3% for politically sensitive figures during the event.

To meet the stringent latency requirements of live‑stream moderation, the multimedia processing cluster extracts a frame every five seconds, stores it in OSS, and pushes a message to the security service. The architecture adopts asynchronous message decoupling, async callbacks, batch processing, and peak‑shaving techniques, reducing node usage by about 70% and cutting average request latency from 3 seconds to 200 ms, with 98% of requests returning within 600 ms.

Overall, the system successfully identified 82 violating streams (68 by algorithm) for pornographic or low‑brow content and over 1,600 streams involving sensitive personalities, while cutting human review effort by more than 90%. The underlying technologies have been packaged into Alibaba Green Net, a content‑risk‑control product offering low‑cost, second‑level response times and >99% accuracy for audio, video, image, and text moderation.