AI is widely used in various industries, corresponding developer community has become rich and varied. They often come from different professional development practice in different scenarios in different areas, which also caused the high cost of learning AI developer of talent, in addition to learning the professional skills needed to understand industry requirements and application scenarios. To this end, initiated Almost Human "AI developer growth plans", joint leader in artificial intelligence to develop common theme courses and projects practices to help developers completed in a shorter period learned from the whole process used.
"AI developers Growth Plan" the first phase of the open class, jointly offered by the Almost Human and micro-public banks, the theme of "FATE federal study entry and application of actual combat," invited partners will also participate in VMware share. Open class for a period of four weeks, a total of six courses, set the theme sharing, project practices, online Q and other sectors, from entry-zero federal study.
Overall learning programs and ways to join, please see: on-line "learning portal FATE and application of federal real" open class!
March 5, at the first lecture "Introduction and application of federal real learning FATE", the micro-public banks Artificial Intelligence Senior Fellow Ma Guoqiang brought "federal study technical presentations, applications and open-source framework FATE" topic to share, detailing the Federal concept learning algorithms, applications, and FATE open source framework introduced and progress, video playback as follows:
[Federal] learning FATE course of the first federal study technical presentations, applications, open-source framework FATE
Lecture featured Q & A:
In the first lecture of the QA session, Ma Guoqiang lecturer answered the little friends are a lot of questions, some of these questions are all widely mentioned, we again made a selection and editing, for your reference.
In addition, we are subject to more questions are also welcome in the comments area of discussion, we will regularly invited lecturer Q & A in the comments area, and continues to be featured questions and answers included in the text of the content of this article.
Question 1: FATE version now supports dynamic joining and leaving it?
FATE is currently not supported by this mechanism, we have previously considered this issue, similar to the application in the mobile terminal. We will continue to assess the feasibility of follow-up and add to our list of requirements inside.
There are plans to support it under win: Question 2?
For now we do not have direct support in the windows program, there are some problems in the use of the package on windows, will pit some additional stability in production is also almost.
Question 3: Transverse federal homomorphic encryption learning is like?
We do not have to use in addition to the lateral transverse federal inside LR homomorphic encryption technology, or homomorphic encryption is not the core, we mainly use the FedAvg and Secure-Aggregation, essentially ideological use of random noise, among participants twenty-two noise, offsetting the server side. Further aspects lateral with less than the homomorphic encryption, as DP (privacy check points) and more usually will mpc programs.
Question 4: secret sharing and SPDZ is used in landscape or portrait in?
Currently used to calculate the longitudinal inside the Pearson correlation coefficient for SPDZ, now, we support two major parties, MPC feeling a little problem on the multi-party programs, including research pysyft when we found that there is a problem in many ways.
Question 5: training a standard model how long?
And a specific algorithm, and the other samples, and data related to amount. In our test environment and test machines, federal and non-federal several times to hundreds of times the gap, but not particularly big difference, because the federal portrait inside a time-consuming mainly homomorphic encryption, lateral Federal , because the use of FedAvg mechanism, in fact, is relatively fast.
Question 6: How much training secureboost slower than revenue?
SecureBoost recommend you use another homomorphic encryption method, affine transformation (IterativeAffine), this is our first contribution of a contributor, this method allows you to have a better experience. In training, it supports more data, training a lot faster, memory consumption is also much less. In our tests, such as stand-alone version, with xgboost difference a few times it is possible. SecureBoost affect the speed of the main features is the dimension of feature dimensions relatively low, but the gap is not so big. Feature dimensions increases, the gradient histogram calculation, addition homomorphic encryption, is time consuming, usually do not, because the encryption is a single number, a lot more time consuming than addition, but when the gradient histogram statistics , time consuming additions will be reflected.
Question 7: How do the characteristics of vertical alignment?
In fact, there are profiles in our share in: through the intersection of privacy (PSI) technology jointly obtained samples. If you want to know more about PSI, Internet search about in our FATE also have some explanation, FATE is the main achievement of RSA and hash of the program, in addition to other programs, such as OT or Bloom Filter .
Question 8: How to learn to deal with the federal data silos?
Federal learn to solve data silos exist. We can see that by sharing just by federal learning algorithm that effect with the effect of the learning center is the same, so this problem does not exist naturally.
Question 9: would consider future technology roadmap with fully homomorphic it?
Fully homomorphic encryption there is a problem of low efficiency. Of course, it is very convenient because the fully homomorphic support a variety of operations, such as addition homomorphic encryption we use, it only supports the addition and several multiply. But the efficiency of the fully homomorphic not go up, so we have not planned this one. Compared with the state and semi-state with full efficiency, there is an order of magnitude difference.
