Auditable Privacy Deep Learning via Blockchain Incentives
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning platforms face challenges in ensuring the confidentiality, auditability, and fairness of parameter sharing among collaborators without a trusted third-party platform, particularly in maintaining the security and reliability of shared parameters during collaborative deep learning processes.
Innovation Solution
A blockchain-empowered incentive mechanism is employed to construct an auditable and privacy-preserving collaborative deep learning platform, utilizing the Paillier encryption algorithm for secure parameter sharing, with a three-layered structure comprising an encryption layer, a blockchain layer, and a training algorithm layer, where parameters are encrypted and stored on a blockchain platform like Corda, ensuring immutability and incentivizing participants through a reward system.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If parameters are shared openly for collaborative deep learning, then training efficiency is improved, but parameter confidentiality and security deteriorate
Solution Approach 1:
The system segments parameter sharing into encrypted portions stored on blockchain and decryption keys held by participants. Parameters are divided into ciphertext components that can be freely shared for training, while the decryption capability remains segmented among authorized participants only.
Solution Approach 2:
Homomorphic encryption serves as an intermediary mechanism that allows collaborative training on encrypted parameters without decryption. The encryption scheme mediates between the need for parameter sharing and confidentiality, enabling computations on ciphertexts that yield encrypted results which can be decrypted by authorized parties.
2Reliability
If a centralized platform is used to manage parameter sharing, then system control and auditability are improved, but system reliability and security deteriorate due to single point of failure
Solution Approach 1:
The system extracts the trust management function from centralized control and distributes it to blockchain nodes through consensus mechanisms. The centralized platform's control function is removed and replaced with decentralized consensus, eliminating the single point of failure while maintaining auditability through immutable blockchain records.
Solution Approach 2:
The system transitions from a single-dimensional centralized control model to a multi-dimensional decentralized architecture where control is distributed across multiple blockchain nodes. This dimensional change allows the system to achieve both control and fault tolerance simultaneously through geometric distribution of trust.
3Reliability
If parameter sharing is made transparent for auditability, then collaboration fairness is improved, but parameter privacy and confidentiality deteriorate
Solution Approach 1:
The system applies different quality properties to different aspects of parameter sharing: transparency and auditability are applied to the metadata and transaction records on blockchain, while confidentiality is applied to the actual parameter values through homomorphic encryption. Each layer has localized quality characteristics suited to its function.
Solution Approach 2:
The system creates asymmetric visibility where blockchain nodes can audit the existence, timing, and validity of parameter shares without being able to view the actual parameter values. The audit trail is asymmetrically designed to reveal process integrity while concealing sensitive data content.
4Reliability
If encryption is applied to protect parameter confidentiality, then parameter security is improved, but computational complexity and processing time deteriorate
Solution Approach 1:
The system applies partial encryption action by encrypting only the parameter shares stored on blockchain while keeping decryption keys separate and managed by participants. This partial application of encryption reduces overall computational complexity compared to full encryption of all training data and models.
Data Source
AI summary
Disclosed is a method for constructing an auditable and privacy-preserving collaborative deep learning platform based on a blockchain-empowered incentive mechanism, which allows trainers of multiple similar models to cooperate for training deep learning models while protecting confidentiality and auditing correctness of shared parameters. The invention has the following technical effects. Firstly, the encryption method used by model trainers protects the confidentiality of sharing parameters; furthermore, the updated parameters are decrypted through the cooperation of all participants, which reduces the possible disclosure of parameters. Secondly, the encrypted parameters are stored in the blockchain, and are only available to participants and authorized miners who are responsible to update parameters. Thirdly, the blockchain-based incentive mechanism guarantees the validity of the parameters, where collaborative trainers need to pay deposit when uploading parameters at the beginning and then the shared parameters can be validated. Concretely, if the parameters are invalid, the deposit would be forfeited.