Authoritative list | NetEase Yidun was selected as an outstanding case of generative AI technology and application by China Academy of Information and Communications Technology

We will further carry out research work, based on generative artificial intelligence technology, deeply explore its application and industrial development status, summarize and promote outstanding results, and promote high-quality development of the industry. In February 2023, the China Academy of Information and Communications Technology (hereinafter referred to as "CAICT") relied on the Generative AI Working Group of the AI ​​Engineering Promotion Committee of the Key Laboratory of Artificial Intelligence and Application Evaluation of the Ministry of Industry and Information Technology to officially launch the Generative AI Working Group. Collection of AI technology and application cases. On the afternoon of May 31, at the Hangzhou General Artificial Intelligence Forum's Large Model Technology Application Sub-forum, the first round of selection results for outstanding cases of generative AI technology and applications were officially released. NetEase Yidun, a subsidiary of NetEase Intelligence Enterprise, was successfully selected as an outstanding case and won the China Credit The Tongyuan Academy issues a certificate.

1. Content risk control is still the top priority for Internet security

In recent years, with the rapid development of the Internet, network data has shown a spurt of growth. Content in various forms, such as text, images, voice, and video, has become an essential part of people's daily lives and work. The diversification of content forms and the significantly lowered threshold for content creation have brought great convenience and efficiency to our lives. However, the content risk control issues caused by this have become increasingly prominent. Digital content security has become a top priority for Internet security.

2. Traditional content risk control faces new challenges

Digital content risk management and control has the characteristics of wide scope and type, fine granularity, frequent confrontation, and diverse needs and standards. In the past, risk control of digital content usually adopted a "post-customization and perception" approach for identification and protection.

· Post-positioning refers to targeted solutions after harmful types or data appear. Timeliness is often post-positioning, and there is no clear ability to perceive and prevent risks in advance.

· Customization refers to a wide range of harmful information types and scopes. Due to the lack of reliable universal capabilities, continuous model customization training is required for different harmful types, and each customization requires building professional field capabilities from scratch, which is time-consuming and costly. high.

· Perception means that the identification of harmful information is often targeted perception, only targeting the harmful type of the target, without making good use of scene content information, that is, past content risk control was based on perceptual level identification rather than cognitive level reasoning. .

·At the same time, there are often many subjective, boundary, and detail differences in the definition of harmful types. Frequently adapting models based on various replacements of standards cannot truly achieve differentiated, hierarchical, and precise protection.

Therefore, how to perceive and prevent possible security risks in advance, how to quickly build security protection capabilities for emerging hazard types, how to comprehensively integrate scene information for in-depth cognitive reasoning, and how to differentiate and hierarchical precise protection have become Important challenges and difficulties in digital content security.

3. AIGC brings new ideas to content risk control

The current development of AIGC makes it possible to provide more universal, front-end, and quick-response digital content risk control capabilities. AIGC injects "world knowledge" to have broader general capabilities, creative capabilities, data perception and knowledge fusion capabilities. Specifically:

· Based on its general capabilities and further customized domain security capabilities, the time cycle and cost of content risk control will be greatly reduced;

· Based on its creative ability, content risk control will sense and prevent unknown risks in advance, transforming "post-position" into "pre-positioning", reducing the hidden dangers of unknown harmful types;

· Based on its rich information injection and fusion capabilities, it fully utilizes and integrates comprehensive information other than harmful types such as scenarios, backgrounds, and knowledge, which will enhance the scenario understanding and knowledge transfer capabilities of content risk control, and carry out deeper cognitive logic. Reasoning and comprehensive prevention and control;

· Based on its prompt context learning paradigm and thinking reasoning process, content risk control will more conveniently adapt to different standards without updating the model, and provide differentiated, hierarchical and layered precise prevention and control.

4. NetEase Yidun: Generative AI technology empowers content risk control

Based on the above ideas, NetEase Yidun uses generative AI technology to develop AIGC-based generation confrontation prevention and control, small sample harmful information identification, fine-grained adaptive identification, comprehensive information logical reasoning and other solutions, built with the strongest spear The strongest shield. details as follows:

(1) Generate confrontation prevention and control plan based on AIGC

In order to achieve early awareness and prevention of security risks, based on AIGC's creative capabilities, NetEase Yidun has developed an AIGC-based confrontation prevention and control plan, improving the original "discovery, deployment and control" to "discovery, generation, deployment and control". This solution will generate harmful types and samples that cannot be covered by the current model or have poor identification effect. Based on the current mainstream security risk types, it will further simulate the security risk change trend through the AIGC method to achieve early awareness of security risks. Further deployment and control methods include combining AIGC-generated samples for joint training of iterative prevention and control models, building AIGC-generated adversarial libraries for fixed-point prevention and control, etc.

(2) AIGC small sample harmful information identification scheme

In order to achieve efficient and rapid construction of digital risk control capabilities and further develop professional field capabilities based on AIGC's general capabilities, we have developed a small sample harmful information identification solution based on AIGC. Improve the original plan of building "general capabilities and domain capabilities" from scratch into the steps of building "AIGC universal capabilities, universal capability compensation and domain capabilities". Through the design of the small sample general ability compensation module, the gap in the application of AIGC general abilities in professional scenarios is bridged at the cost of a small number of samples, and is directly linked to the domain ability building module. Using the method of "AIGC universal capability and universal capability compensation" is faster, more efficient and lower cost than building universal capabilities from scratch.

(3) Fine-grained adaptive recognition scheme based on AIGC

In order to achieve differentiated, hierarchical and precise prevention and control, based on the AIGC prompt context learning paradigm and thinking reasoning capabilities, we developed an AIGC-based fine-grained adaptive identification solution and applied the idea of ​​AIGC prompt learning to the solution of harmful content understanding. In the scheme, and by exploring the alignment of multi-modal prompts and inference inputs with different standards, different standards are mapped through different prompts and inference inputs. For example, in the past, it was difficult to dismantle the sexy genre in a more fine-grained manner, or the cost of dismantling it was very high. Now it can be broken down into more fine-grained dismantling, such as the sexiness of beach scenes and the sexiness of nightclub scenes, thus refining the genre standards. For scenarios and samples, it can better adapt to different subjectivities and different boundary standards, and achieve more fine-grained and hierarchical precise prevention and control.

(4) AIGC comprehensive information logical reasoning solution

In order to achieve cognitive logical reasoning and comprehensive prevention and control, based on AIGC's information injection and fusion and logical reasoning capabilities, we have developed an AIGC-based comprehensive information logical reasoning solution. Improve the original perception recognition that only targets harmful information, so that it can integrate all information including harmful information for cognitive reasoning. AIGC's visual language model is used to obtain comprehensive information except harmful information and used for comprehensive reasoning. The original solution based on "extraction of harmful information and decision-making based on domain capabilities" is improved to a solution based on "extraction of harmful information based on domain capabilities, extraction of general knowledge based on AIGC, information fusion and comprehensive logical decision-making based on visual language models".

5. Generative AI brings significant benefits to content risk control

In digital content risk control scenarios, NetEase Yidun develops and integrates generative AI capabilities to achieve efficient and rapid construction of domain security capabilities, early perception and prevention of security risks, cognitive reasoning and comprehensive prevention and control, and accurate differentiation and grading. Prevention and control. The launch of relevant solutions has helped Yidun fully couple AIGC capabilities from the three levels of capabilities, data, and information, and achieve effects, costs, timeliness, diversity, and adaptability in digital content risk control scenarios actually applied by customers. , significant gains in stability and other aspects.

The above is the detailed content of Authoritative list | NetEase Yidun was selected as an outstanding case of generative AI technology and application by China Academy of Information and Communications Technology. For more information, please follow other related articles on the PHP Chinese website!