Secure Enclave Model Training with Verifiable Digests
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In an untrusting environment, existing technologies face challenges in collaboratively training machine learning models using private data from multiple parties without divulging the data, and ensuring the integrity and verification of the training process.
Innovation Solution
A secure enclave is used within an untrusted host platform to train a model on private data from multiple parties, generating salted hashes and a digest, which is stored on a blockchain, allowing participants to verify the model's training without exposing their data, while ensuring data integrity and security.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a centralized platform is used to store and maintain data, then data management and control are simplified, but data security and integrity cannot be ensured in an untrusting environment with multiple parties
Solution Approach 1:
The system segments data storage and control across multiple decentralized nodes rather than using a single centralized platform. Each participant maintains their own data locally, and the distributed ledger technology segments the trust verification process across multiple independent validators, resolving the contradiction between centralized management ease and distributed security requirements
Solution Approach 2:
The patent introduces cryptographic intermediaries including hash functions, digital signatures, and encrypted commitments as mediators between parties. These cryptographic mechanisms enable verification of data integrity and authenticity without requiring direct trust between participants, allowing secure collaboration in untrusting environments while maintaining operational simplicity
2Measurement precision
If multiple parties collaborate to train a machine learning model using private data, then model accuracy improves, but data leakage and security risks increase
Solution Approach 1:
The system creates cryptographic copies of data through hash commitments and encrypted representations. Participants work with these cryptographic copies rather than raw private data during the training process, allowing model training to proceed with improved accuracy from multiple data sources while the original private data remains protected and cannot be reconstructed from the copies
Solution Approach 2:
Cryptographic intermediaries including secure multi-party computation protocols and homomorphic encryption serve as mediators that enable collaborative model training. These intermediaries allow computations to be performed on encrypted data or through coordinated cryptographic protocols, ensuring that no single party can access another's private data while still contributing to an accurate aggregated model
3Reliability
If data is encrypted and stored on a blockchain, then data integrity and verification are ensured, but system complexity increases
Solution Approach 1:
The patent employs universal cryptographic primitives such as hash functions and digital signatures that serve multiple functions simultaneously. These same cryptographic mechanisms provide data integrity verification, authentication, and tamper detection across different layers of the system, reducing overall complexity by using a unified cryptographic approach rather than separate mechanisms for each function
Solution Approach 2:
The system extracts and separates the verification functionality from the storage functionality. Cryptographic hash commitments and digital signatures are extracted as independent verification layers that operate separately from the underlying storage mechanism, allowing the system to ensure data integrity through verifiable cryptographic proofs without complicating the core storage and retrieval operations
4Reliability
If a secure enclave is used to train models on private data, then data security is improved, but verification of the training process becomes difficult
Solution Approach 1:
The system implements feedback mechanisms where the secure enclave generates cryptographic proofs of training execution that are returned to participants. These proofs include verifiable computation results, hash commitments of training data, and cryptographic signatures that provide feedback confirming the training process occurred as expected, enabling verification while maintaining the security benefits of the enclave
Solution Approach 2:
The system performs preliminary cryptographic setup actions before the actual training process. Participants pre-generate cryptographic keys, establish commitment schemes, and set up verification protocols in advance. These preliminary actions create a verifiable framework that allows the secure enclave to train models confidentially while enabling subsequent verification of the training process through pre-established cryptographic checks
Data Source
AI summary
An example operation may include one or more of receiving, from a plurality of providers by a secure enclave, a plurality of training data sets and a plurality of salts paired with the plurality of training data sets, respectively, hashing, via the secure enclave, pairs of training data sets and salts to generate a plurality of salted hashes, and combining the plurality of salted hashes to generate a digest, training, via the secure enclave, a model based on the plurality of training data sets to generate a trained model, and providing the trained model and the digest to a plurality of data providers of the plurality of training data sets and the plurality of salts.