Neural Information Processing Systems 2021
(NeurIPS 2021) Demos
Overview of TripleBlind Toolset
Blind Learning and Privophy Demo
Some Preliminary Results
Blind Learning
Dataset:
CIFAR-10
Model:
ResNet-18
Training Criteria:
Accuracy per training round over 100 global epochs. The baseline is a model trained on the complete dataset (50,000 images) at a centralized location for 100 epochs, using ResNet-18. The baseline accuracy is 80%. The visualized results also show the training results on three individual clients, each with 16,600 non-overlapping images for both IID and Non-IID data distributions.
Results:
Overall, BL and FL algorithms can provide acceptable utility compared to centralized training. However, FL requires more training rounds and is more communication and computation costly compared to BL. The following figures illustrate the accuracy/round for BL, FL, centrally-trained model, and individual models per client.
Test accuracy per training round for non-IID data distribution over 3 clients. E: local epochs.
Test accuracy per training round for IID data distribution over 3 clients.
Privophy
Dataset:
MNIST and CIFAR-10
Models:
- LeNet-5
- VGG-16
- Network A — 3 fully connected layer with ReLU
- Network B — 1 Convolutional layer, 2 fully connected layers with ReLU
- Network C — 2 Convolutional layers, 1 fully connected layer with ReLU
Training Criteria:
Compare total execution time to current state-of-the-art protocols, including SecureNN, Gazelle, MiniONN, Chameleon.
Results:
See the figure below, times in seconds.
Comparison of the supported layers and functions
Interactive Demo Plan
Introduction
We’ll discuss the motivation behind TripleBlind and what it means to “unlock private data sharing.”
Explanation
We’ll cover the underlying methodology of our innovations, including Blind Learning and Privophy.
Use Case
We’ll demo the use of TripleBlind with a live example where we train an image classifier using two remote decentralized datasets.
Hands On
We’ll invite the audience to play with our solution.
Audience Interaction Plan
Members of the audience will be invited to interact with our system. Each participant will play the role of either a data scientist or a data owner.
Data Scientist
The data scientist task will focus on training a deep learning model using remote decentralized datasets.
Data Owner
The data owner task will focus on running an encrypted inference using a remote model on some local data.
Accessible through live Jupyter Notebooks
Each notebook will include one of five scenarios (see below). Note that each dataset is divided into two sets, each set is placed on an individual Google Cloud instance to simulate training and inference on decentralized data owned by different organizations.
Training a deep learning model (VGG-16) for image classification using CIFAR-10
Training a deep learning model for tabular data classification
Training a deep learning model for multi-modal (text and images) classification
Running a secure inference task using a remote, pre-trained model
We have also added another notebook to illustrate our Private Set Intersection protocol
Note:
The Jupyter notebooks server will be kept alive between the periods 08/27 to 10/07 for the reviewers to access and test the notebooks. However, we will reduce the computational capacity for these servers in this period due to their cost. We will improve the computational capacity for the server during the actual demo to accommodate as many participants as possible.