About Triple Blind
TripleBlind is a provider of data and algorithm security services.
Overview of Briefing
0:00 – 2:00
Das reviews the Triple Blind platform and a hypothetical use case involving Capital One and Experian discussing how the data for issuing a credit card is typically exchanged by multiple parties. Challenges due to GCPR, CCPA, leftover data that is no longer needed, and the solution Triple Blind has developed is significantly faster than traditional methods while protecting access to the data used to train a model.
2:00 – 4:20
Discussion of Triple Blind methods for sharing access to encrypted datasets and additional applications. A problem statement was presented that different banks have different types of data that they would like to use to predict a future transaction.
Using real data a neural network is trained during the demo to come up with the prediction of future spend. 3 datasets were used while maintaining privacy, preprocessed the data, to train the deep neural network
4:20 – 7:00
Using multiple browser windows the solution is demonstrated and different datasets shown and discussed. Data shown also demonstrates the ability for Triple Blind to analyze different data sets sold in the portal. Specifically, a NASA dataset was shown as an example.
Demonstration follows retrieving the data, preprocessing, training the model, and running the neural network model to predict the next likely transaction.
7:00 – 9:45
Script is run live demonstrating finding the dataset and the requested permissions displayed so the third party must grant access to the underlying dataset. Once permissions are granted the deep neural network is trained after the cryptographic consent is in place. This method allows the data to be used only for this specific purpose. The trained object is privacy maintaining.
9:45 – 12:05
Discussion of how the algorithm can be kept separate from the data as is needed in healthcare environments. Triple Blind differentiators were shared including the ability to train a model in minutes instead of weeks.