We are excited to announce that our tool demonstration paper has been accepted to NeurIPS 2021, Demo Track. The paper is entitled: “TripleBlind: A Privacy-Preserving Framework for Decentralized Data and Algorithms.”
The paper sheds light on the trilemma of accuracy-privacy-communication challenges for training and inference using decentralized data and algorithms. In particular, we demonstrate two major innovations of TripleBlind: Blind Learning and Privophy. Blind Learning is our training paradigm for neural networks. Privophy is a set of cryptographic protocols that enable efficient privacy-preserving processes; e.g., secure inference-as-a-service. Both innovations enable AI practitioners from all domains (i.e., physicians) to run real-time decentralized training and inference using multi-institutional datasets via a set of semi-automated, easy-to-use APIs.
Our demonstration will focus on three major concerns when it comes to privacy-preserving AI using decentralized data and algorithms: accuracy, privacy, and communication cost. We illustrate how Blind Learning can match the accuracy of centrally trained models without having to transfer any raw data outside the owner’s organization while being more communication efficient than rival solutions, including Federated Learning and Split Learning. We also demonstrate in real-time the efficiency and ease of use of our secure multi-party computation inference protocol.
The audience will also get a chance to interact with our toolset via Jupyter notebooks. The notebooks will be placed on cloud compute nodes that play the role of realistic organizations with datasets and models.
We invite you to join our demo session on Friday, Dec. 10 at 11:20-11:35 a.m. CT
For more information, visit TripleBlind’s webpage at the NeurIPS website. You can also read our tool demo proposal.
The audience can use our APIs to train deep models using these datasets and also perform secure inference in real-time. We will provide five types of notebooks during our demo presentation that allow the audience to try the following tasks:
- Training an image classifier using CIFAR-10 images distributed over three clients
- Training a classifier using tabular data distributed over two clients
- Training a multi-modal classifier using data of different types (text and images) distributed over two clients
- A private set intersection to find common IDs between two clients using secure MPC
- A secure MPC inference using a model and data owned by two different parties without compromising the privacy of neither the model nor the data