I'm interested in whether federated learning can bring ML to situations where you don't want to pool all the data in one spot for privacy reasons. Say you run a B2B Saas business in seperate tenancies and in each tenant contains sensitive information about that client (eg about their clients). Could you run a federated learning model such that it could learn in each tenant and improve the overall model but not share any of the sensitive information between tenants?
This depends on details of the ML model. There is a mathematical field devoted to this specific question, differential privacy, and the techniques are in production at scale at Google, Apple and in the US census.
Yes indeed. Using federated learning with Flower makes training over multiple disconnected partitions possible. Additionally there are privacy enhancing techniques such as differential privacy but they come with a cost and are not always nesesecary.
You can train models over multiple silos, devices, users and many other kind of partitioning where for some reason you can't aggregate the dataset centrally.
Comments
I'm interested in whether federated learning can bring ML to situations where you don't want to pool all the data in one spot for privacy reasons. Say you run a B2B Saas business in seperate tenancies and in each tenant contains sensitive information about that client (eg about their clients). Could you run a federated learning model such that it could learn in each tenant and improve the overall model but not share any of the sensitive information between tenants?
This depends on details of the ML model. There is a mathematical field devoted to this specific question, differential privacy, and the techniques are in production at scale at Google, Apple and in the US census.
Yes indeed. Using federated learning with Flower makes training over multiple disconnected partitions possible. Additionally there are privacy enhancing techniques such as differential privacy but they come with a cost and are not always nesesecary.
You can train models over multiple silos, devices, users and many other kind of partitioning where for some reason you can't aggregate the dataset centrally.