Table 2.

Summarized Taxonomy for Federated Learning Systems.

Taxonomy	Category	Structure	Advantages
Data partitioning	Horizontal FL	Different parties and similar data features	Holds larger variety of parties
	Vertical FL	Similar parties and different data features	Holds larger variety of data features
	Federated transfer learning	Different parties and different data features	Holds larger variety of parties and data features
Machine learning models	Linear models	Linear regression, ride regression, lasso regression	Ease of implementation
	Decision tree	gradient boosting, decision trees, random forests	Accurate, stable, and can map non-linear relationships
	Neural networks	-	Learning capabilities, highly robust and fault-tolerant
Privacy mechanisms	Model aggregation	Central manager learns by aggregating the locally trained model	Avoid transmitting original data
	Cryptographic methods	Using encryption algorithms such as homomorphic encryption and secure multi party computation (SMC) to encrypt the messages exchanged among parties	Enables the calculation and processing of encrypted data
	Differential privacy	Reducing the impact of a single data record on the calculation of the global model	Reduce the effect of data poisoning attacks
Methods for solving heterogeneity	Asynchronous communication sampling	To resolve the heterogeneity of parties	Solve the problem of communication delays and avoid simultaneous training with heterogeneity of parties
	Fault-tolerant mechanism	To resolve the failure of parties	Prevent whole system from collapsing if one of the parties failed
	Heterogeneous model	To resolve the heterogeneity of data	Resolve the issue of models diversity
Communication architecture	Centralized design	Architecture controlled by a central aggregation manager/server	Reduces communication cost
	Decentralized design	Communication performed among parties without the existence of a central manager/server	Reduces the risk of backdoor attacks
Scale of federation	Cross-silo FL	Parties are less in number, hold large amounts of data and equipped with high computation power	Fits for FL among institutions
	Cross-device FL	Parties are high in number, hold less amount of data and equipped with less computation power	Fits for FL among individuals
Motivation of federation	Regulations	Motivated by laws such as GDPR and others
	Incentives	Motivated by desire of updating some services	Enhancing ML services