FedML & Blockchain
A part of our platform for privacy-preserving federated machine learning using blockchain technology and smart contracts with support from Vinnova, Sweden's innovation agency.
Industry and expert advisory panel:
Trust and Security in FedML
When working on a federated machine learning model in a setting with several different actors, there is a challenge to trust that the model generated is secure, maintain full data privacy and is not misused by anyone in the group. One of our FedML projects is about integrating blockchain technologies in our FedML platform.

Working computational schemes for FedML are already available in the public domain, and one key to wide commercial adoption is in building a highly distributed, secure and reliable platform to allow scalable alliances among participants that don't fully trust each other. This problem has the following key challenges:

  1. Enabling trust in the quality and integrity of the produced federated ML model despite lack of trust between alliance members.
  2. Automated evaluation of terms and conditions governing the alliance during computation.
Figure 1
Overview of the federated learning alliance where alliance members train on local datasets and contribute to a global machine learning model while actions are logged on an immutable blockchain, and with smart contracts for governance.
Our Solution
We are addressing these challenges using blockchain and smart contract technology. Specifically, we are:

  1. Using blockchain to create an immutable audit trail for federated models for greater trustworthiness in tracking and proving provenance.
  2. Enhancing encryption and communication strategies between nodes and federated model to maintain better privacy-preservation.
  3. Demonstrating the use of smart contracts for governing the business logic of a FedML alliance.

Figure 1 illustrates the envisioned solution. The blocks in the chain contains the references to parameters of ML models, and with each new update of the global model, a new block will be added in the chain. These immutable blocks contains the evolution of a model from the beginning to its final stage. Smart contracts are used to define rules and penalties around an agreement in the same way that a traditional contract does, but also automatically enforce those obligations. In the context of FedML, we are using smart contracts to define the rules for the model training. They will also be used to restrict the number of operations, enable different incentives based on the member's contributions and to define new rules consisting of upcoming requirements. The use of smart contracts will encourage alliance members to contribute independently and earn incentives based on their contributions. It will also help to mitigate errors. The proposed approach of using smart contracts for rule-based model training is unique and has not been tested before within the domain of federated machine learning.
Strategic Relevance
The proposed platform can address the needs of a range of organizations working with sensitive data. For example due to the new European data protection regulations (GDPR), companies need to take extra measures for data management and security. The approach of FedML aligns well with the requirements from the GDPR. It allows data owners to have complete control, yet it is possible to apply advanced analytics on the datasets. At the end of the project we will arrive at a platform for federated machine learning alliances with the following capabilities:

  1. Transparency and auditability of the federated machine learning process by enabling access to immutable audit trails
  2. A smart contract based system for rule setting and enforcement in formation, model training and access to federated machine learning models
  3. A toolset to build markets and incentives around formation, training and access to federated machine learning models
Conclusion
Our solution will contribute to increased trust in the digital society by allowing advanced machine learning to be developed on distributed data with full respect of the confidentiality of the data providers and the property rights of the companies that propose the machine learning models. It will create opportunities for both small and large scale organizations to work together and securely build highly accurate models that was not possible before due to the strict data sharing policies.
The platform is being built with the support from Vinnova, Sweden's innovation agency.
We will post project updates regularly on our blog. Let us know if you are interested in exploring privacy-preserving federated machine learning (FedML) using blockchain technology and smart contracts.
Phone: +46(0)18-7770303
E-mail: contact@scaleoutsystems.com

We are located at Epicenter in Stockholm and Juvelen in Uppsala.