Maximilian Wisgickl

Advisors
Arne Küderle (M.Sc.), Prof. Dr. Björn Eskofier

Duration
02/2020 – 08/2020

Abstract
With the ongoing digitalization of health data around the world, the opportunity arises to use machine learning to extract knowledge from this ever-growing pool of data [7]. The application of machine learning (ML) on big data has been shown to increase the accuracy with which diseases are detected and has the potential to support disease prediction and prevention [3]. Traditionally, data needs to be gathered in a central place to apply machine learning algorithms. However, healthcare data is of a sensitive personal nature and needs protection against mischief. If it is stored and processed, a provider needs to comply with a multitude of regulations to address potential danger to privacy and security. These regulations also need to be followed by ML-researchers and developers, which severely complicates their work, in terms of bureaucratic overhead and adjustments to implemented software to be compliant. Therefore, new approaches to use medical data are currently under development. One of them is federated learning, which poses a valid alternative solution to centralized data storage. In this strategy, only the model itself persists centrally, but its training takes place directly on the user’s device, negating the need to collect personal data at a central location. Each user who wants to contribute to the enhancement of the model receives a version of the central model, trains it on their device, and returns the result. The global model aggregates all user-improved models until its accuracy is sufficient. This allows the model to use data from a theoretically endless amount of users without the need to collect data and hence infringe user’s privacy. This concept has already been used successfully outside of the medical field. The most prominent example is the use of federated learning for a next-word prediction algorithm of smartphone keyboards that would otherwise need to send the private message data to a server [5]. However, currently available frameworks that could perform federated learning require a specific programming language and runtime [2]. This restriction is acceptable for the industrial context. In research, however, we want to be free to implement algorithms in languages and runtimes that best fit the topic and are suitable for rapid prototyping of new concepts. In the proposed work, it is planned to solve this problem by creating an infrastructure that poses fewer restrictions on language or runtime as well as eases the step of ML-deployment, which is currently complicated [1]. For example, in deployment, a model is usually continuously improved. To be able to do this, data and models are repeatedly validated, and if they yield improvements over previous versions, an infrastructure that distributes the enhanced version is needed. The proposed infrastructure would allow scientists to prototype their algorithms on distributed networks without requiring much technical knowledge and hopefully lead to more projects using data that would have been unavailable in a non-federated context. In addition, it is planned to create a matching service that allows researchers to distribute the algorithms to specific users who possess the required data. This should further reduce the effort of deploying a distributed algorithm. To achieve this, we will expand the existing ONE infrastructure [6]. ONE already implements a data storage mechanism together with a data distribution protocol and auditing of all data and events, inherently supporting features needed for regulatory compliance. The contribution will consist of a runtime abstraction, an API for ML-distribution, and a matching service to match researches to patients who possess the data needed for their ML-algorithms. The runtime abstraction will make use of Docker [4], a reproducible virtualization layer that allows executing code in an isolated and predefined environment. The overall goal is to create a distributed, ML-ecosystem with very few restrictions to language or runtime environment in addition to supporting the research data “collection” process. The application and concepts will undergo performance and usability evaluation by a focus group consisting of ML-researchers. The usability evaluation will investigate how easy the infrastructure can be integrated into existing ML-projects and small test projects. The performance tests will contain a run of a simple federated evaluation algorithm to determine functionality and comparison in its execution time, comparing a centralized and distributed approach. The group feedback will dictate an iterative improvement cycle.

References:
[1] Denis Baylor, Eric Breck, Heng-Tze Cheng, Noah Fiedel, Chuan Yu Foo, Zakaria Haque, Salem Haykal, Mustafa Ispir, Vihan Jain, Levent Koc, Chiu Yuen Koo, Lukasz Lew, Clemens Mewald, Akshay Naresh Modi, Neoklis Polyzotis, Sukriti Ramesh, Sudip Roy, Steven Euijong Whang, Martin Wicke, Jarek Wilkiewicz, Xin Zhang, and Martin Zinkevich. TFX: A TensorFlow- Based Production-Scale Machine Learning Platform. In KDD 2017 Applied Data Science Paper, pages 1387–1395, Halifax, NS, Canada, 2017.
[2] Keith Bonawitz, Hubert Eichner, Wolfgang Grieskamp, Dzmitry Huba, Alex Ingerman, Vladimir Ivanov, Chloé Kiddon, Jakub Konecn´ Konecn´y, Stefano Mazzocchi, H Brendan Mcmahan, Timon Van Overveldt, David Petrou, Daniel Ramage, and Jason Roselander. Towards Federated Learning at Scale: System Design. Technical report, 2019. [3] Thomas Davenport and Ravi Kalakota. The potential for artificial intelligence in healthcare. Technical Report 2, 2019.
[4] Docker. What is a Container? App Containerization Docker. https://www.docker.com/resources/what-container.
[5] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated Learning: Challenges, Methods, and Future Directions. aug 2019.
[6] REFINIO. Refinio. https://refinio.net/.
[7] Kangkang Wang, Rajiv Mathews, Chloé Kiddon, Hubert Eichner, Françoise Beaufays, and Daniel Ramage. Federated Evaluation of On-device Personalization. http://arxiv.org/abs/1910.10252, 2019.