CaliFusion

Your Data, Your Control
Collective Intelligence, Powered by ICP and Calimero.

This Project won the first prize in the Calimero x ICP hackathon main track

A platform where data providers can share encrypted datasets for AI model training. ICP smart contracts ensure fair payments, and Calimero preserves data confidentiality.

Read our wonderful docs here.

Pitch deck here.

Key Features and Architecture

Overview

This architecture describes a system for model training and publishing, integrating three main components: a Backend service, Calimero infrastructure, and ICP (Internet Computer Protocol) for secure token management.

System Components

Backend

A distributed system that serves as the central coordination point, handling:

• Key exchange protocols
• Model verification
• User authentication
• Communication between different system components

Calimero

Two separate Calimero infrastructures are implemented:

1. Model/key storage and management
2. User-specific data processing and model deployment

ICP (Internet Computer Protocol)

Handles cross-contract calls for token transfers and payment verification, ensuring secure financial transactions within the system.

User Types and Interactions

User A (Model Publisher)

• Participates in Diffie-Hellman key exchange with the Backend for AES encryption
• Uploads encrypted models and corresponding keys to Calimero
• Receives aggregated values and reports from the system

User B (Model Trainer)

• Models are securely fetched to user's local machine
• Training occurs in the Model/Model.py file
• Interacts with Backend for model verification

Security Features

1. Key Exchange

• Implements Diffie-Hellman protocol
• AES encryption for model security
• Secure key distribution system

2. Backend Verification

• Ensures proper key exchange
• Verifies reference commitments
• Manages system integrity

3. Payment Processing

• Cross-contract calls via ICP
• Token-based access control
• Secure payment verification

Technical Implementation

Model Publishing Flow

1. Publisher initiates Diffie-Hellman key exchange
2. Backend validates and processes encryption
3. Model and keys are uploaded to Calimero
4. System confirms successful publication

Model Training Process

1. Trainer receives encrypted model
2. Local processing on trainer's machine
3. Results verification through Backend
4. Integration with Calimero infrastructure

Access Control

1. Users make token payments
2. ICP verifies transactions
3. Access granted to Model key
4. Local model decryption enabled

Notes

This architecture prioritizes security, scalability, and user privacy while main