Welcome to visit netease Cloud Community to learn more about Netease’s technical product operation experience.


On May 19, LiveVideoStack and netease Cloud Shield jointly created the “Entertainment multimedia development and application Practice” topic, to help developers and operators of pan-entertainment platform, improve technical ability, break through difficulties, expand ideas and vision.


In special BBS, netease cloud easy to shield the CTO Nicholas has Shared the AI fu to the content of the safety technology practice topic content, from three aspects of policy, technology, products, netease was introduced in detail in content security system construction of practice experience, including how to establish fine clear of harmful information audit system and emergency response mechanism, How to adopt intelligent recognition technology based on natural language and deep learning technology to ensure comprehensive and accurate coverage of text, pictures, voice, especially live video content and so on?


Current state of the content security industry: Harmful information continues to increase


In 2017, there were 3.4 billion Internet users worldwide, with an Internet penetration rate of 46%. There are 2.8 billion smartphones in the world; In 2016, the Internet accounted for 55% of Chinese users’ media time; Mobile Internet usage exceeds TV…… These figures show that the cash ability of network broadcast has far exceeded games, TELEVISION and other media.


Zhu haoqi said that while the Internet has further penetrated into every corner of people’s lives, it has also brought about a continuous increase in harmful information — the number of illegal and undesirable information reported remains high, and has even attracted national strategic attention.


Without cyber security, there can be no national security, still less should the Internet become a platform for spreading harmful information and spreading rumors to cause trouble. Therefore, in recent years, all kinds of punishment cases are too numerous to enumerate.


The netease Cloud yi Shield CTO listed: April 18, 2017, Beijing Network information Office, public Security Bureau cultural field administrative law enforcement team joint interview toutiao, volcano live, pepper live; At the end of April 2017, The State Administration of Press, Publication, Radio, Film and Television ordered Tencent to rectify the broadcasting of audio-visual programs on its wechat official account and mobile client. On June 22, 2017, sarFT ordered the closure of video services on Sina Weibo, AcFun and Ifeng.com. In July 2017, a number of Bilibili episodes were pulled from the shelves; 6 April 2018; In order to cooperate with the rectification of SARFT, I quickly expanded my internal audit team; On April 8, 2018, sarFT ordered Toutiao to permanently shut down vulgar audio-visual products such as “content jokes”…… It can be said that the industry has entered a strong regulatory era.


“For enterprises, content security has become a matter of life and death.” For example, in April, documents released by the Ministry of Culture and the Ministry of Tourism showed that 4,939 live-streaming apps had been screened and 370 removed, Zhu said. “Therefore, many livestreaming app platforms have chosen to increase censorship efforts and labor input.” For example, Kuaishou posted a recruitment showing that it would expand the number of auditors to 5,000 with a monthly salary of 4,000-9,000 yuan. Toutiao also said that it will expand its review team to 6,000 people in the near future, and expects to reach 10,000 people……





Zhu Haoqi believes that the content security of Internet live broadcasting platform brings more challenges to enterprises than that. In terms of cost, effect and technology, various challenges are constantly upgrading, which brings great difficulties to enterprises.



How does netease Cloud Easy Shield provide security protection at the technical level?


After sharing the grim situation, Zhu Haoqi talked about how to do content security of netease Cloud Yi Shield with 20 years of technology precipitation.


He said, Easy shield in content security, a total of three stages:


  • Stage 1: Keywords, blacklist and whitelist, filter;

  • Stage 2: Content feature recognition (skin color, texture), Bayesian filtering, similarity matching, rule system;

  • Stage 3: Big data analysis (user behavior, user classification), human-machine recognition, artificial intelligence (semantic recognition, image recognition, natural language processing, etc.).


What are the core technologies of text, image, voice and video? Zhu Haoqi said there are mainly the following:


  • Text: Intelligent keywords, fuzzy matching, semantic analysis, small language detection;

  • Image: image recognition, feature recognition, image retrieval, OCR recognition;

  • Speech: MD5 library detection, speech feature retrieval, speech feature recognition, speech recognition;

  • Video: video transcoding compression, video stream analysis, screenshot/frame analysis.


The subsoil in the technical line of the CTO, said thanks to the rapid development of the neural network subject, in recent years, machine learning technology is also in constant progress, for content security provides the most solid technical support, including a layered biological neural networks, image recognition technology, the depth of neural network, the neurons activated heat map, etc., in the security technology gives more reinforcement effect.





Hubel& Wiesel (Image: Harvard Brain)


These technologies, including image recognition, audio filtering and video filtering, are also fully used in netease Yidun, he said.





  • Image recognition technology: support portrait recognition, support pornographic image detection, support animation recognition; Support similar matching, support picture rotation, cropping, support to change picture brightness, tone; Supports watermark identification.

  • Audio filtering technology: Based on the powerful voice recognition and analysis engine, netease will filter harmful voice in real time through the garbage database accumulated for many years, and then convert it into text for further filtering. The key technical points are the combination of speech to text and text detection.

  • Video filtering technology (live broadcast, on-demand) : video stream detection, video screenshot/frame detection, heat monitoring in live broadcast room, anchor blacklist, machine + manual. The technical point lies in the close combination of video conversion image and image recognition technology.



For netease, the diversity of internal data also provides yi Shield with rich practice scenarios. “Among the hundreds of thousands of variations of data built by netease Mailbox, netease News, Cloud music, LOFTER, cloud Album, Mint Live and other internal applications, Yidun can realize real-time encryption and filtering of spam, security and brush prevention and other functions, and maintain the operating atmosphere and user experience of the product well.”


The real technical environment is more complex, how does netease YIDun go further?


In fact, from the perspective of technology, manpower, cost and other dimensions, content security protection in the real environment is more complex. For example, assuming that one out of 10K images is a prohibited image, the accuracy of ESHIELD system is 99.9%, that is, 1K images will judge a wrong image, then the calculated False Positive Rate (that is, the proportion of normal samples in the collection of prohibited images) is:





That is, 10 of the 11 results judged to be prohibited are misjudged — the data is difficult to interpret, hackers’ continuous attacks have put forward higher requirements for content security protection.


  • Indecipherable: As shown in the figure below, an image of a correctly recognized giant panda was identified as a gibbon with 99.3% confidence after adding tiny changes that were barely legible to the human eye.







  • Hacker attack: By changing the dimensional characteristics of the input data, it can attack the data processing link of deep learning, resulting in escape attack and data pollution attack on applications such as deep learning image recognition. For example, a star in the picture below can be recognized by the machine as “a certain mountain” after the data is processed in a dimensional way.







Not to mention, content security “attackers” have already formed a mature industrial chain. They are groups on the fringes of the law, whose interests are clear and coordinated, and whose means of attack continue to escalate. Therefore, for all content security vendors, this is a high cost, long see-saw battle.


Zhu Haoqi said, “Based on this, Compared with the traditional single-dimension content recognition service in the industry, YIDun also provides multi-dimensional technical services of account number, IP, behavioral action (follow, vote, comment), content recognition, user portrait, group portrait, environmental data and device fingerprint. In addition, it is in-depth analysis of attack vulnerabilities, from registration protection, login protection, environmental monitoring, behavior analysis, user analysis, offline computing of the whole service process for in-depth security protection.”






At present, based on the strong user relationship network built by netease, YIDun provides scenario-oriented solutions. From advertising filtering, smart as yellow, violence could identify, rumors the four aspects, the barrage real-time filtering, chat text/image filtering, image/nickname filter, live broadcast TV wall monitoring, studio audio, anchor blacklist, user list and room temperature analysis comprehensive security protection function. In addition, the operation mode of “machine + manual” is more subtle and provides upgraded technical functions, such as intelligent machine monitoring, priority push, intelligent sorting, suspect highlighting, warning support, cut-off operation, automatic filling, video size adaptation and so on.







Heavy technical defense, the safety effect is self-evident. Zhu haoqi shared that At present, E-Shield provides security for hundreds of millions of netease users and content data every day, and also provides comprehensive information security services for all netease products. “E-shield’s comprehensive information security services have provided significant protection to netease’s products in the report of illegal and undesirable information on major commercial websites in March this year.” He said.


How to use easy shield quickly?


Netease Yidun is a fast and intelligent content security cloud service, focusing on content security solutions. At present, netease Yidun has achieved penetration application in pan-entertainment, multimedia, operators, e-commerce, financial securities and other industries. Under the coverage of customers in multiple fields, it is a firm and unbroken security technology guarantee with original intention.


Content security data visualization, intelligent live TV wall, battle-hardened audit team, rich functional applications, now available.



Netease eSHIELD provides you with anti-spam and anti-ddos security services. Please click here for free trial.


From poster vision to web design