Privacy-by-Design Collaborative Machine Learning

Data owners have access to valuable information which demands data privacy, for example, banks maintain depositors’ financial records, hospitals keep patients’ healthcare data, and an ISP holds its infrastructure information. All such data is highly sensitive, and any data leaks can have a detrimental impact on individuals and businesses. 

In addition, there are privacy regulations across the world, like the General Data Protection Regulation (GDPR) in Europe and the California Consumer Privacy Act (CCPA) in the US, that legally govern the sharing of private data and enforce the control and rights over such data.

Model owners design sophisticated models, however, often lack data from diverse sources to train generalizable models. Model owners demand model confidentiality, for example, a pharmaceutical company designs a model to help new drug designs, a FinTech company develops a fraud detection model to catch financial anomalies, or a solution provider designs a model to help ISPs identify infrastructure issues. 

Models are becoming innovative, and model owners spend significant time and resources to develop and train their models. Thus, both the models and training methods are often intellectual properties and need to be kept confidential.

The widely adopted industry solutions to collaborative machine learning today include Federated Learning and Split Learning. Here, a model is trained across multiple data owners holding local datasets, without exchanging them. However, such solutions have acknowledged limitations regarding the strong protection of both data privacy and model confidentiality.

We have created the Privacy-by-Design Collaborative Machine Learning service that protects both data privacy and model confidentiality throughout the entire ML training pipeline. Our service ensures each data owner’s dataset is treated in a provable privacy-preserving way. In addition, the confidentiality of model owners’ models and training methods is strongly protected where no entity other than the model owners ever has access to confidential models, even in a trustless environment.

With our service, there is no need for AI/ML developers or practitioners to change their daily routines. Instead, they program their model training code and then invoke our service APIs or use our GUI in the same way as they would program it for Federated Learning; thus, providing an easy-to-adopt, non-breaking experience for today’s developers and practitioners.