Under the current background, data islands are an important obstacle restricting the development of AI. There are many data barriers within the company and between departments of subsidiaries. Barriers between different institutions are even more of a problem. At the same time, privacy exposure or data leakage is also a very serious social problem, and it is likely to become a "hardest hit area" of supervision. Facebook, for example, was hit with huge fines in the EU for data breaches. Professor Li Xiaolin said: Eliminating data islands in various industries and innovating models to enable data collaboration are the future trends, and in between, knowledge federation can play an important role.

Li Xiaolin emphasized: Knowledge Federation is not a single technical method, but a set of theoretical framework system. Isolated islands protect data privacy while utilizing data. The goal of the Knowledge Federation is to create a data-safe artificial intelligence ecosystem, effectively utilize multi-party data through data security exchange protocols, and carry out knowledge co-creation, sharing, and reasoning, so that data is available and invisible.

The knowledge federation structure is divided into four layers, information layer, model layer, cognitive layer and knowledge layer. The information layer includes technical security queries. The model layer includes many models, such as linear models, tree models, some commonly used financial models and deep learning models. The cognitive layer includes transfer learning, cross-media expression and grassroots learning.

Li Xiaolin pointed out that the core concept of knowledge federation is "data is usable but not visible, knowledge can be created and shared". MIT once said in an article that "according to the update of the gradient, the original data can be reversed." But this problem is in the knowledge federation model. does not exist in . The Knowledge Federation is aggregated on the gradient of the ciphertext space, so it is safe. In the future, Tongdun Technology also hopes to create a standardized exchange protocol with colleagues from all walks of life.

secondary title

How does knowledge federation apply to actual scenarios?

Talking about the scenario application of knowledge federation, Professor Li Xiaolin took financial anti-fraud, epidemic forecast, AIDS prevention and control as examples, and mainly introduced: knowledge federation can break the data island, extract the cognitive layer, and after the cognitive layer is extracted , and then do some integration with the third party through encryption. By using some migration methods or original agreement methods, the data barriers of various institutions can be broken to achieve the effect of joint defense and joint control, which can be applied to anti-fraud or corporate credit investigation. For example, a small and micro enterprise has very limited information, so it may have some related transactions or affiliated companies, the business owner and his social relationship, social network, social activities, etc. These information can be used in series. These knowledge graphs Combined, you can more effectively judge the credit situation of the enterprise.