Federated Machine Learning
Federated Machine Learning is an approach to train machine learning models through decentralized data sources. As opposed to centrally trained models, no data aggregation is necessary.
This solves many privacy and data security concerns, and thus enables many new use cases in privacy sensitive domains such as finance, IoT and proprietary data.
How Fed ML works
In Federated Machine Learning, there are several participants, each holding their own private dataset.
A trusted coordinator orchestrates the AI training by selecting participants and asking them to refine (or “train”) the AI model using their own private data. They only share the result of this training step with the coordinator and not their actual training data itself.
The full AI training process consists of multiple training rounds and works as follows:
The platform manages the global master model that is fed by all individual models.
The XAIN coordinator algorithm selects a few participants to join the next round of training.
The platform pushes the global model to all selected participants.
Each selected participant receives a copy of the latest, most efficient global master model. This provides higher accuracy to the next round of training.
They then follow with a training round based on that AI model, using only their local private data.
The coordinator algorithm recollects the latest learnings of all AI models from each participant.
The updates get aggregated to form a new improved AI model. We use secure dynamic aggregation to ensure the highest quality mix of models.
Advantages of Federated ML
Bringing AI capabilities to SMEs
Our technology will allow an easier onboarding of smaller companies that may neither have AI capabilities in-house nor datasets that are rich enough to optimally support their decision processes.
Federated Machine Learning allows companies to train machine learning models whilst also benefiting from training based on other companies’ datasets. This approach will dramatically improve companies’ decision support based on AI without having to rely on dominant AI companies and their business models.
Prof. Michael Huth
Professor of Computer Science at Imperial College London & CTO of XAIN
Diagnosis software in healthcare
Diagnosis algorithms need a lot of case data to learn from.
- Patient data is highly privacy relevant.
- With the XAIN technology, models can be trained based on decentralized data buckets.
Workflow automation software (ANDY)
Enterprise data is often siloed in.
- Learn across data silos while keeping the data as is.
- Improve model performance through decentralized learning.
Customer service bots
SaaS software providers have to train individual client models.
- Train a consolidated model without compromising data privacy regulations.
- Pretrained models offer less training time to onboard new software customers.