The FeatureCloud Platform for Federated Learning in Biomedicine: Unified Approach

Abstract

Background: Machine learning and artificial intelligence have shown promising results in many areas and are driven by the increasing amount of available data. However, these data are often distributed across different institutions and cannot be easily shared owing to strict privacy regulations. Federated learning (FL) allows the training of distributed machine learning models without sharing sensitive data. In addition, the implementation is time-consuming and requires advanced programming skills and complex technical infrastructures.

Objective: Various tools and frameworks have been developed to simplify the development of FL algorithms and provide the necessary technical infrastructure. Although there are many high-quality frameworks, most focus only on a single application case or method. To our knowledge, there are no generic frameworks, meaning that the existing solutions are restricted to a particular type of algorithm or application field. Furthermore, most of these frameworks provide an application programming interface that needs programming knowledge. There is no collection of ready-to-use FL algorithms that are extendable and allow users (eg, researchers) without programming knowledge to apply FL. A central FL platform for both FL algorithm developers and users does not exist. This study aimed to address this gap and make FL available to everyone by developing FeatureCloud, an all-in-one platform for FL in biomedicine and beyond.

Methods: The FeatureCloud platform consists of 3 main components: a global frontend, a global backend, and a local controller. Our platform uses a Docker to separate the local acting components of the platform from the sensitive data systems. We evaluated our platform using 4 different algorithms on 5 data sets for both accuracy and runtime.

Results: FeatureCloud removes the complexity of distributed systems for developers and end users by providing a comprehensive platform for executing multi-institutional FL analyses and implementing FL algorithms. Through its integrated artificial intelligence store, federated algorithms can easily be published and reused by the community. To secure sensitive raw data, FeatureCloud supports privacy-enhancing technologies to secure the shared local models and assures high standards in data privacy to comply with the strict General Data Protection Regulation. Our evaluation shows that applications developed in FeatureCloud can produce highly similar results compared with centralized approaches and scale well for an increasing number of participating sites.

Conclusions: FeatureCloud provides a ready-to-use platform that integrates the development and execution of FL algorithms while reducing the complexity to a minimum and removing the hurdles of federated infrastructure. Thus, we believe that it has the potential to greatly increase the accessibility of privacy-preserving and distributed data analyses in biomedicine and beyond.

Keywords: AI store; additive secret sharing; artificial intelligence; federated learning; interactive platform; privacy-enhancing technologies; privacy-preserving machine learning.

Figure 1

Outline of the FeatureCloud system. Medical institutions collaborate in a federated study with all primary or raw data remaining at their original location. FeatureCloud handles the distribution, execution, and communication of certified artificial intelligence (AI) applications from the FeatureCloud AI store and addresses developers and end users.

Figure 2

System architecture of FeatureCloud with 2 participants. The controller, frontend, Docker engine, and application instances run locally at each participant’s site. The FeatureCloud backend and Docker registry are running on FeatureCloud servers. The relay server can be run on a separate server, or participants can use a provided instance from FeatureCloud. The components are connected via transmission control protocol/IP connections (straight lines). All links are http based, except for link 7, which uses a raw socket connection. Links 1 to 3 use JSON for serialization, and links 4 to 6 use the Docker application programming interface.

Figure 3

Workflow structure used for evaluation. The first application (purple—Cross-Validation) creates splits for cross-validation (CV). All following applications perform their tasks on each split individually, in a federated fashion, only transmitting model parameters. The gray dots represent intermediate training and test data. The second application (green—“Normalization”) performs normalization, and the third application (blue—“Random Forest”) trains the models, generating a global model based on the output of the normalization application. The resulting global model is evaluated in the evaluation application (orange—“Evaluation [Classification]”). The evaluation results are finally aggregated to obtain an evaluation report based on the initial CV splits.

Figure 4

Performance evaluation of federated artificial intelligence methods. The box plots show the results of a 10-fold cross-valuation for the different classification and regression models and data sets in multiple settings. Only the deep learning model was evaluated on a test set. The centralized results are shown in orange, the corresponding federated results in blue, and the individual results obtained locally at each participant in gray. Each model was evaluated on the entire test set (dark gray) such as the centralized and federated models and on the individual (local) parts of the test set (light gray). The federated logistic and linear regressions perform in identical fashion to their centralized versions, and the federated random forest and deep learning models perform in similar fashion to their centralized versions. BRCA: Breast Invasive Carcinoma; ILDP: Indian Liver Patient Dataset; SHARE: Survey of Health, Aging and Retirement in Europe.

Figure 5

Runtime and network traffic. The left plots show runtime for unlimited and throttled connections, the right plots show network traffic for the coordinator and participants evaluated on the ILPD. The lines represent the median values measured over 10 executions. The areas show the 25% and 75% quartiles to illustrate variance across the executions. ILPD: Indian Liver Patient Dataset; s:second; B: byte; M: million.