Privacy computing is a key technology for the trusted circulation of data elements. It is an important direction of current academic and industrial research and a key measure to help the development of my country's digital economy. In order to promote the development of privacy computing and provide a communication platform for people in the privacy computing industry, on July 26, the 2023 Privacy Computing Conference jointly sponsored by the China Academy of Information and Communications Technology and the China Society of Communications was successfully held in Chengyang District, Qingdao, attracting more than 10,000 people. focus on.
Impulse Online, as the vice-chairman unit of the Privacy Computing Alliance, was invited to participate in this event. Founder & CEO Liu Yao delivered a speech titled "Application Practice of Privacy Computing in the AI Big Model Era", sharing that in the era of large models, Impulse Online thinking, judgment and best application practices. With the rapid development of large model technology, China is also showing a booming scene in this field. From the open source bottom-level large model framework to upper-level applications, domestic companies have invested in the research, development and application of large model technology. As a privacy computing technology company, Impulse Online has also been paying attention to the relationship between large models and data privacy protection in recent years.
Since large models have a high dependence on the amount of training data. For example, Open AI uses tens of billions or even trillions of data sets to train large models. Therefore, in the era of AIGC large models, data privacy and security protection are increasingly important. highlight. On July 13, the Cyberspace Administration of China also issued the "Interim Measures for the Management of Generative Artificial Intelligence Services", which clearly requires large model service providers to protect the privacy of users' input information and usage records, and to regulate the prediction of generative AI. Responsible for the legality of the sources of training data and optimization data.
As a result, privacy computing technology and AI computing power are deeply integrated, and the construction of "privacy-enhanced AI computing power" has become the core infrastructure. From the perspective of AIGC industrial structure, it can be divided into three layers. The bottom layer is the pre-training model with huge amounts of data built by Open AI, Google, etc.; the middle layer is fine-tuning based on the pre-trained models in different industries to generate models in fields such as finance and government affairs; the top layer is the diversified application scenarios generated by the middle-layer model, such as Language generation, image generation, etc. These three layers are relatively hierarchical and decoupled, and each layer can be combined with privacy computing technology to play an important role. In the pre-training of the underlying model, Trusted Execution Environment (TEE) technology can be used to protect the data security and privacy of the training, inference, and fine-tuning processes. The mid-level industry model is being fine-tuned, and a data circulation and trading platform can be built to facilitate secondary training. In top-level model applications, privacy technology can be used to protect model security reasoning and orchestration.
It can be seen that Trusted Execution Environment Technology (TEE) based on the combination of software and hardware can better support the versatility of AI. It can not only support AI models of different frameworks, but also support the training and prediction of large models. It can also ensure the security of models and data in GPU accelerated operations through TEE direct-to-GPU technology.
Since its establishment, Impulse Online has always insisted on in-depth development of trusted execution environment technology that combines software and hardware. It has supported compatible adaptation of all domestic CPU chips, and was the first to release the first batch of privacy computing all-in-one computers and heterogeneous privacy computing all-in-one computers. Since 2021, it has begun to explore running large models, deep learning and other AI algorithms in the Trusted Execution Environment (TEE) to protect the security and privacy of data. Recently, Impulse Online has joined forces with Haiguang, Kunpeng, Tianshu Zhixin, Kunlun Core and other GPU chip manufacturers to launch the trusted AIGC all-in-one machine. As the industry's first solution that can benchmark NVIDIA's confidential computing power GPU, in the era of large models, It has successfully transplanted mainstream open source model capabilities to the TEE environment and supports protected training and inference, ensuring that user data interacting with large models is not lost or leaked, and the inference process is safe and controllable.
From a low-level perspective, Impulse Online introduces CPU Trusted Execution Environment (TEE) technology to GPU to protect the safe training and inference of machine learning and deep learning. It also opens up new application scenarios for domestic GPU manufacturers.
At the chip and large model level, "model as a service" has become a mainstream business model, and in this model, security and privacy protection are particularly important. As the middle layer, Impulse Online helps protect privacy and security throughout the entire process of large model training, prediction, and even commercial applications through in-depth cooperation with underlying chip manufacturers.
Privacy computing helps banks build trusted AI model SaaS service platforms
Impulse Online built an AI model SaaS service platform based on a trusted execution environment for a bank, migrating various large model capabilities such as OCR, NLP, graphics, and language to confidential computing nodes, and opening services to the outside world through APIs to achieve Models are shared openly instead of being deployed privately multiple times. For example, in a contract comparison scenario, when the user needs to analyze and process the contract through the platform, the contract can be transmitted to the privacy computing TEE node using encryption, achieving no contact, no storage, and secure computing of user data. , effectively solves the privacy and security issues faced in model SaaS services. Help banks achieve open sharing of models, and also allow more institutions to obtain credible AI services.
02
Privacy computing helps China Telecom build a "privacy and security trust computing center"
Impulse Online cooperated with China Telecom to build a secure and private computing platform based on a trusted execution environment in its intelligent computing center. At the same time, the project won the "Top Ten Benchmark Cases of Privacy Computing" in the 2022 Galaxy Case. This solution deploys OCR, NLP, text generation, image processing and other large model reasoning capabilities in the intelligent computing center, and allows third-party users to use these models safely. User data is directly transmitted and encrypted and processed in the TEE environment to avoid being collected and stored by the center. It not only protects user data privacy, but also enables the intelligent computing center to achieve neutrality and credibility as a computing service provider. Telecom can rely on the cloud to output self-developed AI model capabilities, reduce costs, increase efficiency, and improve computing power utilization efficiency.
unveiling event
Release of results
Part
1
At the conference, the China Academy of Information and Communications Technology, together with many units including Impulse Online, jointly released the "Trusted Digital Network (TDN) Test Bed", which set up more than ten items in terms of resource access, circulation calculation, circulation control and interconnection. Inspection and testing capabilities, and the construction of some key capabilities has been completed. In the future, on the one hand, the test bed will serve the access, review and monitoring during the construction and operation of the trusted digital network. On the other hand, it will also be able to provide a series of customized tests for more technology providers and application demanders in the industry. Services to help enterprises improve data circulation technology, products and application capabilities.
Part
2
At the conference, the China Academy of Information and Communications Technology and the Privacy Computing Alliance jointly released the "Privacy Computing Application Research Report (2023)" with industry units to analyze how privacy computing empowers the circulation application of data elements. Impulse Online participated deeply in this event as the vice-chairman unit of the Privacy Computing Alliance Report writing helps the high-quality development of the data element market. !