Triple Blind
Triple Blind Technical Briefing
Published January 19, 2021
Supported by Shopify , The Kauffman Foundation and Levvel
Overview
TripleBlind demos a cryptographic platform for privacy-preserving data and algorithm exchange — an alternative to the status-quo model of decrypting and replicating third-party data (the Capital One / Experian problem). TripleBlind\'s novel one-way encryption is homomorphic-like but roughly a trillion times faster than full homomorphic encryption, and it works across text, voice, video, and images. The data can only be used for the purpose explicitly authorized, and algorithms themselves can also be encrypted — protecting IP and training data from reverse-engineering. The live demo uses Kaggle\'s customer-transaction-prediction dataset split across three mock organizations (Standard Chartered, JPMorgan Chase, BNP Paribas) running in three different browsers. Privacy-preserving EDA lets a data scientist see shape, distribution, and plots without seeing individual records. A feed-forward neural network (dense + ReLU layers) is trained across all three parties — each granting cryptographic consent with a justification — and produces a privacy-preserving PyTorch object that can be used for local inference or secure multi-party compute predictions (both algorithm and data stay hidden from each other). Training runs in minutes where FHE would take weeks, and the code is drop-in compatible with PyTorch, Pandas, TensorFlow, and XGBoost — GDPR, CCPA, and FDIC compliant.
0:00 Introduction to TripleBlind
TripleBlind enforces privacy cryptographically across every data-and-algorithm exchange. Where Capital One today decrypts Experian credit files and creates replicas inside its own systems (creating GDPR and privacy risks), TripleBlind lets parties operate on data while it stays encrypted.
1:00 Novel one-way encryption — trillion× faster than FHE
TripleBlind has built a new kind of encryption that works on text, voice, video, or images — homomorphic-like in spirit but roughly a trillion times faster than full homomorphic encryption. Data can only ever be used for the purpose explicitly authorized.
1:30 Algorithms can also be encrypted
Not just the data — the algorithm itself can be encrypted, protecting the IP inside the model and the training data from reverse-engineering attacks. This is zero-trust for vendor/third-party data partnerships.
2:30 Kaggle-based customer transaction prediction
The demo uses a public Kaggle customer transaction prediction dataset. The problem: banks want to use each other's data to improve predictions, but sharing under CCPA/GDPR and competitive concerns forces most teams to use synthetic data — which loses accuracy.
4:00 Three-organization data fusion across browsers
Three browser windows (Edge, Brave, Chrome) represent three organizations. The demo shows bringing datasets together under TripleBlind — preserving privacy across all three parties while still training on real data.
4:30 Data discovery with privacy-preserving EDA
A data scientist can explore size, shape, layout, and distribution of a dataset via privacy-preserving EDA — viewing plots and descriptive statistics that never leak individual records in the clear.
6:30 Training a deep neural network across parties
A feed-forward neural network (dense + ReLU layers, loss function, target column) is trained across the three mocked banks (Standard Chartered, JPMorgan Chase, BNP Paribas). Each org grants cryptographic consent with a justification before the join occurs.
8:00 Trained PyTorch object — local and MPC inference
The output is a PyTorch-trained object that preserves privacy — no reverse engineering is possible. Predictions can run locally against data that never leaves the user, or via secure multi-party compute where both the algorithm and the data stay hidden from each other.
10:30 Speed, PyTorch compatibility, and compliance
The key differentiators: homomorphic-grade privacy in minutes (not weeks), drop-in compatibility with PyTorch / Pandas / TensorFlow / XGBoost, GDPR/CCPA/FDIC-compliant data sharing, and reduced bias by accessing data previously locked behind privacy concerns.
Presented by Riddhiman Das — Triple Blind · LinkedIn
