IEEE 3652.1-2020 pdf download

IEEE 3652.1-2020 pdf download.IEEE Guide for Architectural Framework and Application of Federated Machine Learning.
The module has the following interface:
INPUT: A list of all sample identifiers in different data sources
OUTPUT: A list of all overlapped sample identifiers
7.3.2 Feature alignment module
A feature alignment module is mainly used for horizontal federated machine learning. The module identifies the overlapped features of different data sources and does not disclose non-overlapping features and sample IDs.
The module has the following interface:
INPUT: A list of all feature names in different data sources
OUTPUT: A list of overlapped feature names in different data sources
7.3.3 Federated feature engineering module
A federated feature engineering module can be used for horizontal and vertical federated machine learning, as well as federated transfer learning. The module identifies overlapped input features and determines a set of new features that are relevant to the learning task.
The module has the following interface:
IN PUT: The data set that is used for feature engineering
— A list of all overlapped features in data sources
A list of identifiers of all overlapped samples in data sources
OUTPUT: A updated set of features of data sources
Notice that the outcome of feature engineering is dependent on the learning task. Since different applications of federated machine learning algorithms have different requirements for features, this module should he customized according to specific requirements e.g., for time-domain features, space domain features, and frequency domain features, etc.
7.3.4 Federated machine learning algorithm module
A federated machine learning algorithm module covers all kinds of basic algorithms needed by federated machine learning for different application scenarios or tasks. It should include, but not be limited to, the following algorithms:
A set of federated machine learning algorithms that can be used in supervised learning, semi-supervised learning, and unsupervised learning scenarios, e.g., tree-based algorithms, deep learning algorithms [B22], linear learning algorithms of federated machine learning.
The module has the following interface:
IN PUT: the data sets that are used for federated machine learning
A list of all overlapped features in the data sets, a list of all overlapped samples in the data sets, a set of hyper—parameters controlling the learning algorithm
— OUTPUT: (part of) the learned FML model
Notice that there is a diversity of FML models, such as trees and neural networks, that should be supported by the federated machine learning algorithm module.
7.3.5 Algorithm evaluation module
An algorithm evaluation module should evaluate federated machine learning models according to various evaluation measures. Refer to Clause 9 for details. The module should include, but not be limited to, the following measures:
Performance evaluation
Efficiency evaluation
Privacy-preserving evaluation
Security evaluation
The module has the following interface:
IN PUT: The learned FML model
The testing data set that is used to evaluate the FML model
A list of all overlapped features in the testing data set
OUTPUT: A list of evaluation results (performance, efficiency, security and privacy, etc.)
7.3.6 Contribution evaluation module
A contribution evaluation module should evaluate the contribution to the overall performance of the FML model provided by each owner’s raw data.
The module has the following interfiuce:
INPUT: The training data sets from all data owners
The testing data set that is used to evaluate the contributions
A list of all overlapped features in the data sets
A list of all overlapped samples in the training data sets
The hyper-parameters controlling the learning algorithm
The used federated machine learning algorithm
Output: A list of scores indicating the contributions of all data owners
Notice that the outcome of this module should assist in interpreting the decision-making logic of the FML model and indicate the contributions made by each owner’s raw data. The outcome should also facilitate the investigation of the root causes of underlying problems.