• 

  Front End

• 

  AWS EC2

• 

  Solidity Contract

• 

  Zero knowledge proof circuit (circom)

About the Project

Our project focuses on creating a decentralized federated learning system for hospitals, ensuring privacy and data confidentiality while enabling collaborative AI training across different institutions. The system allows hospitals to train AI models on their own medical data (like chest X-rays), and then securely share model parameters with a global server that combines and averages these models without ever accessing the sensitive data itself.

What inspired us to work on this project is the growing need for privacy-preserving machine learning, especially in the healthcare industry. With increasing concerns about data security and strict regulations like HIPAA, it’s critical to find ways to leverage data without violating confidentiality. Federated learning, combined with blockchain and cryptography techniques like zero-knowledge proofs (ZKPs) and CKKS encryption, offers a solution to these challenges.

Through this project, I learned how decentralized systems can foster innovation and collaboration without compromising data security. We integrated technologies such as Ethereum-based smart contracts and privacy-preserving cryptography to ensure model integrity and correctness, even without accessing the data itself.

How We Built the Project

Data Privacy and Model Training: We designed two isolated hospital environments where each hospital trains a CNN model on its dataset. Federated Server and Blockchain: Each hospital sends its model’s weights and training metadata (loss function, optimizer, model architecture) to a global server, which uses CKKS encryption to perform privacy-preserving weight averaging. Zero Knowledge Proofs: To verify the training integrity without revealing sensitive data, ZKPs are employed, allowing the server to confirm the validity of training without accessing the actual data. Smart Contracts for Verification: The verification process is handled using Ethereum-based smart contracts written in solidity and built over Rust-based Circom circuits, that ensure the authenticity of the training logs. Frontend for Transparency: The frontend displays two key components: (1) Information about the global model, including its type, optimizer, and weights, and (2) Decentralized training logs, showing client IDs and training information.

Challenges Faced One major challenge was implementing privacy-preserving technologies like CKKS encryption and ZKPs into the federated learning workflow while maintaining the efficiency of the system. Another was designing the smart contract to handle training log verification and ensuring secure communication between hospitals and the global server. We also had to overcome resource constraints, particularly in handling large model weights in a decentralized and encrypted manner.

Built With

• amazon-web-services
• css
• django
• flask
• html
• javascript
• python
• solidity
• tensorflow
• typescript