FLChain

`We have built an AppChain for the reliability of federated learning data and learning using Cosmos-SDK. Additionally, federated learning is conducted off-chain through the flwr framework.

A. AppChain (built with CosmosSDK) Features (Modules) include Record, Token (Mint, Transfer, Lock, UnLock), Event, and ZK Prove.

Record : Registers Data Info (Hash of Data, MetaData(features, count, memo)) and records learning results.
Token : An AppChain Token used to pay for learning. When applying for learning, a certain amount of Token is Locked. After learning is completed, participants - prove their learning and UnLock (Transfer) their tokens.
Event : Emits an Event for the start of federated learning.
ZK Prove : A Prove Module that verifies honest participation in learning with the recorded Data. It uses Raw Data as a Private Input and Data Hash, w (pre-learning weight), and a (learning rate) as Public Input. The Output is w' (post-learning weight).

Entities include DataProvider and DataConsumer (learning applicants).

The sequence is as follows:

The DataProvider records its data with a Hash value and associated metadata (information describing the data).
The DataConsumer queries the Chain for data information and finds the required DataProvider.
The DataConsumer applies for learning, sending a designated DataProvider List, Model URI, and their verification key (for the ZK Verifier).
Chain Nodes that receive this Tx (Token) from step 3 validate the Token, Lock it, and then emit an Event.

B. FlowerFramework (Off-chain)

DataProviders that were listening for Events will set up configurations for federated learning once they listen to the event from step 4. 6.The DataConsumer and DataProviders proceed with learning. The reasons for building our AppChain for data and trust resolution are revealed: -> The authenticity of the RawData used for learning is secured by comparing it with the Hash value registered on the Chain. -> During learning, DataProviders create a ZK Proof of the learning process using their data and send it to the Chain.
The ZK Verifier on the Chain leaves a record of whether the Proof is Valid/Invalid.
After the learning is completed, the participating DataProviders apply for Unlock, distribute the Token to the Valid DataProviders, and return the remaining Tokens to the DataConsumer.

This structure encourages both DataProviders and DataConsumers to actively participate without any disruptive behavior in the learning process.

Inspiration

Recognizing the need to enhance personal data protection and address the issues within existing systems such as unauthorized data collection, indiscriminate data usage, and security vulnerabilities arising from data trading among companies, our team embarked on the development of the FLChain project. This project is based on Federated Learning (FL) technology and integrates it with the Celestia Chain to establish a secure and efficient data processing platform while building a new app chain.

What it does

FLChain combines Federated Learning with the Celestia Chain to reinforce personal data protection and data security. This project accomplishes the following: Performs federated learning without centralizing data, ensuring personal data protection. Utilizes Celestia Modular + Rollups for fast information processing and reduced transaction costs. Establishes a federated learning-based chain to enhance data protection and privacy, creating a secure and efficient app chain.

How we built it

Implemented data processing and model training based on Federated Learning. Enhanced data processing and security by integrating with the Celestia Chain. Utilized Celestia Modular and Rollups to create a secure and efficient app chain

Challenges we ran into

Learning New Technologies: Understanding and implementing new technologies like Federated Learning and the Celestia Chain presented a learning curve.
Security Enhancement: Overcoming security challenges to ensure user data protection and privacy. Data Processing and Performance Optimization: Addressing difficulties related to processing large datasets and optimizing system performance.

Accomplishments that we're proud of

Integration of Federated Learning and Celestia Chain: Effectively integrating Federated Learning and the Celestia Chain to create a secure data processing environment.
Data Protection (FL + ZK prove): Safely managing and protecting user data through enhanced personal data protection and security measures.

What we learned

Acquiring New Technical Skills: Improving our ability to acquire and implement new technologies. Security and Privacy: Recognizing the importance of understanding and implementing security and privacy measures for user data protection.

What's next for FLChain

Adding More Features: Incorporating additional data protection and security features to enhance user trust.
Research and Improvement: Continuously researching and improving the project, focusing on Federated Learning and the Celestia Chain to provide better performance and security.