IEEE 3652.1-2020 pdf free download – IEEE Guide for Architectural Framework and Applicaton of Federated Machine Learning.
4. FML reference architecture
Federated machine learning is a distributed machine-learning framework that enables multiple participants to collaboratively train and use a machine learning model for a given task, e.g., classification, prediction, and recommendation. Within this framework, all raw data owned by different participants are protected by secure and privacy-preserving techniques, which prevent the data from being tampered and disclosed by other participants or reverse-engineered by other participants. An FML framework consists of data, users, and systems that are illustrated in Figure 1.
In the framework, data are distributed across different repositories and are used to build FML sub-models that are, subsequently, integrated to result in a federated model in a secure and privacy-preserving manner.
An FML user can be a natural person, a corporation, or other organizations with the legal capacity to participate in the FML framework. In addition, FML users may play four roles, which are the data owners, the model users, the coordinators, and the auditors (see details in Clause 6). The FML system consists of multiple functional modules that provide FML services to users (see more detail in Clause 7).
5. FML data view
5.1 Overview FML data is often stored in a standard database format, i.e., tables, whereby each row represents a data sample and each column represents a feature or label of the sample. In supervised learning, a complete training data set consists of both features, denoted by X, and labels, denoted by Y. A set of feature attributes are usually represented as a feature vector (X1, X2, …, Xn). Moreover, a unique sample ID is associated with each data sample. In an FML system, data from multiple data sets may have overlaps in sample IDs and/or feature attributes. Depending on the extent of overlapping along either sample IDs or features, the following three cases are of interest for a federated machine learning framework:
— The overlap of feature attributes (X1, X2, …) is substantially larger than the overlap of sample IDs (U1, U2, …)
— The overlap of sample IDs (U1, U2, …) is substantially larger than the overlap of feature attributes (X1, X2, …)
— The overlap of sample IDs (U1, U2, …) and the overlap of features (X1, X2, …) are both small
In the list above, “substantially larger” (and “signifcant overlap”) is judged by whether the overlapping data can be used to build a high-quality machine learning model, where the measure of quality is determined by applications.
Depending on the application scenarios, FML is categorized as Horizontal FML, Vertical FML, and Federated Transfer Learning (as shown in Table 1). Also, depending on whether the data sets used have the same format or different format, FML is categorized into homogeneous FML or heterogeneous FML.
5.2 Horizontal FML Horizontal FML refers to building a model in the scenario where data sets have significant overlaps on the feature spaces (X1, X2, …) but not on the ID spaces. In this case, an FML model can be built as if the data is split and join horizontally. Horizontal FML may apply to scenarios where the number of sample IDs from data owners is too small to build a high-quality model. An FML model should perform better than the sub-models built by one single data set, and the performance of the FML model is very close to that of the model built when all data were put together in one location.IEEE 3652.1 pdf download.