Neurocrypt

Neurocrypt is a privacy-preserving AI solution developed during the Privacy Preserving AI Hackathon in partnership with Entrepreneur First, Hugging Face, and Zama. The project focuses on secure processing of sensitive EEG brain wave data using Fully Homomorphic Encryption (FHE), allowing real-time analysis while maintaining data privacy.

🚀 Hackathon Overview

The Privacy Preserving AI Hackathon, held from September 26-28 in Paris, brought together 50 selected participants out of hundreds of applications. Organized by Zama, Hugging Face, and Entrepreneur First, the event aimed to advance privacy-preserving technologies, with a focus on FHE. Participants were challenged to build innovative AI solutions using Concrete-ML, an open-source library designed for FHE.

🧠 The Problem

EEG (Electroencephalography) measures electrical activity in the brain and is used in:

• Seizure detection
• Brain-computer interfaces (BCIs)
• Mental health monitoring

Despite its potential, EEG data poses privacy concerns, and the devices used to measure brain waves are often resource-limited. Neurocrypt addresses this issue by securely processing encrypted EEG data on the cloud without compromising privacy or performance.

🔒 The Solution

Neurocrypt's approach involves:

1. Encrypting EEG data using FHE: Data collected by wearable devices is encrypted before transmission, ensuring privacy.
2. Cloud-based encrypted machine learning: Leveraging FHE on cloud infrastructure to run machine learning models on encrypted data.

This solution allows for accurate and secure real-time analysis, addressing privacy concerns while still providing valuable insights.

📊 Dataset

We used the Seizure EEG Dataset hosted on Hugging Face, featuring:

• Image size: 224x224 pixels
• Total records: 1,318,793 rows
• Data windows: Captured over 6-second intervals across different brain wave frequencies

The dataset supports our efforts in building and validating EEG-based AI models for privacy-preserving applications.

📈 Model Performance

Non-Encrypted Model

• Achieved 85% validation accuracy for 224x224 resolution images.
• Data was downsampled from 500k to 15k for optimized training.

Encrypted Model

• Post-training encryption was implemented using Concrete-ML, reaching 77% validation accuracy.
• Encryption and processing were executed within 7 minutes, showcasing FHE's potential for real-world applications.

🎥 Demo

The Neurocrypt demo illustrates the end-to-end workflow, from EEG data encryption on a wearable device to FHE-encrypted machine learning analysis on the cloud.

neurocrypt_demo.mov

🏅 Team Members

• Simon Coessens
• Arijit Samal
• Thomas Chardonnens
• Anand-Arnaud Pajaniradjane
• Batu Ergun

📜 License

This project is open-sourced under the MIT License.

🔧 Getting Started

Follow these steps to run Neurocrypt:

1. Clone the repository:

   git clone https://github.com/yourusername/neurocrypt.git
   cd neurocrypt