Federated Learning Privacy via Blockchain and Multi-Party Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional federated learning architectures face challenges such as loss of privacy due to sharing of local models, lack of traceability of participation, and loss of confidentiality of both the algorithm and the data used for training.
Innovation Solution
A computer-implemented method utilizing a blockchain network and multi-party computation (MPC) peer-to-peer network to train algorithms, ensuring confidentiality and privacy by keeping local training results private and not sharing them with the algorithm provider.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If local models are shared to train the global model in conventional federated learning, then the model training can be performed across multiple decentralized devices, but privacy is lost because private information can be extracted from local models
Solution Approach 1:
The patent introduces homomorphic encryption as an intermediary mechanism that allows local models to be processed (added/aggregated) in encrypted form without revealing the underlying private information. The encryption scheme acts as a mediator that enables the federated learning function while preserving privacy, resolving the contradiction between model training capability and privacy protection.
2Loss of information
If conventional federated learning is used to train algorithms with decentralized data, then data privacy can be protected, but traceability of participation is lost leading to repudiation and compensation problems
Solution Approach 1:
The patent implements continuous traceability through blockchain technology that records each participant's contribution to the federated learning process. The system continuously logs model updates, participant identities, and training metrics on the blockchain, ensuring that the useful action of traceability persists throughout the entire federated learning lifecycle, thereby preventing repudiation and enabling compensation.
3Productivity
If the trained model is shared in conventional federated learning, then the algorithm can be obtained for service provision, but confidentiality of both the algorithm and the training data is lost
Solution Approach 1:
The patent employs secure multi-party computation to create functional copies of the trained model that can be used for service provision without revealing the underlying confidential algorithm or training data. Instead of sharing the actual model parameters, the system generates encrypted copies or derived models that maintain confidentiality while enabling productive service deployment.
Data Source
AI summary
A computer implemented method for training algorithms, including providing a data owner having datasets, an algorithm provider having algorithms to be trained and a blockchain network in whereby a training management provider is deployed; sending a training request to the training management provider, including a training manifest; sending the manifest to the data owner; sending by each data owner, an acceptance having a component ensuring the identity of the participating data owner; sending, by the training management provider and once minimal requirements have been met, the acceptance manifests to the algorithm provider; generating, by the algorithm provider, a compiled multi-party computation code; sending the algorithm to be trained and the compiled multi-party computation code to the data owners participating in training the algorithm; training; executing the multi-party computation code generating a multi-party peer-to-peer network; and generating a model based on the local models and the algorithm to be trained.