How NetEase Optimizes AI for Digital Content Risk Control and Cost Efficiency

Conference Overview

On November 25‑26, 2022, QCon Global Software Development Conference was held in Shanghai. NetEase’s Intelligent Enterprise Technical Committee organized a special session titled “Exploring Integrated Communication Technology and AI Commercialization Practice.” The session featured algorithm expert Li Yuke, server development expert Cao Jiajun, audio‑video engine expert Zhu Mingliang, and NLP expert Feng Minwei. This article records Dr. Li Yuke’s presentation.

Speaker Introduction

Li Yuke, an algorithm engineer at NetEase Intelligent Enterprise, works on digital content risk control. He shared how his team improves AI production efficiency in this domain.

Digital Content Risk Control Overview

NetEase’s Yidun Business Center provides AI detection capabilities for text, image, video, and audio, identifying harmful content and offering combined security solutions such as business risk control and mobile security. Services are used internally and by many external enterprises.

Presentation Structure

Business background of digital content risk control.

Cost‑optimization efforts focusing on data, algorithms, and services.

Current and future interesting projects.

AI Business Application

AI is essential for comprehensive content moderation because the internet contains diverse media types. High coverage, strict effectiveness, and security requirements drive the need for robust AI models.

Production Materials

The team has built data collection pipelines, a custom annotation platform, stable labeling teams, extensive compute resources, cloud‑based and domestic‑chip collaborations, and a commercial‑focused algorithm team with customized solutions.

Cost Challenges

As business volume and algorithm modules grow, AI deployment costs rise sharply, including labor, resources, and hardware expenses.

Data Acquisition

Sensitive data is scarce, so the team uses active learning to select valuable samples for labeling. They also employ multimodal collection (e.g., OCR + ASR verification) and cross‑modal retrieval to generate seed data.

Semi‑Supervised Learning

Unlabeled data is leveraged via semi‑supervised methods. Early stages used two‑stage approaches; later, online semi‑supervised training combined labeling and model updates. A dual‑network teacher‑student scheme improves learning by exploiting differences between network types.

Data Selection Strategies

To balance cost and efficiency, the team selects data for manual labeling based on model confidence, feature similarity, and domain diversity, ensuring that hard or out‑of‑domain samples receive human attention.

Algorithm Design

A lightweight front‑end network filters out normal data (70‑80% of inputs). Remaining samples undergo a detection network that extracts candidate regions (patches) for downstream fine‑grained models. Dynamic inference adjusts network depth and resolution based on sample difficulty.

Dynamic Inference and Weakly Supervised Segmentation

The system uses a two‑branch architecture: a classification branch for whole‑image decisions and a segmentation branch that generates coarse masks to guide a second‑stage patch classification, enabling a “coarse‑to‑fine” workflow.

Training Strategies

Joint training of detection and classification networks ensures consistent data flow. Dynamic batch processing and multi‑instance parallelism improve throughput, while dynamic gradient updates focus shallow layers on easy samples and deep layers on hard ones.

Service Deployment

Model acceleration techniques (model pruning, distillation) and open‑source inference engines (TVM, MNN) are employed. The service architecture separates tasks (OCR, face, weapon detection) into independent services, each with decoding, AI inference, and post‑processing stages. Dynamic batching and multi‑instance execution further boost performance.

Future Work

Planned directions include supporting dynamic batch sizes and dimensions, expanding multimodal solutions while controlling cost, and exploring knowledge‑fusion techniques to combine outputs from multiple services into unified models.

Conclusion

The presentation highlighted how NetEase leverages AI to enhance digital content risk control, focusing on data pipelines, algorithmic innovations, and service‑level optimizations to achieve higher efficiency and lower operational costs.