Cryptographically Secure Machine Learning Model Sharing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The practical sharing of machine learning models is difficult due to concerns over the disclosure of proprietary information and the high costs associated with developing and training these models, making it challenging for organizations to share or use trained models without revealing internal details.
Innovation Solution
The implementation of a cryptographically secure machine learning system that employs multi-party computation protocols to select and utilize machine learning models while keeping internal components encrypted, allowing organizations to share or use trained models without exposing sensitive information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If machine learning models are shared or distributed between organizations, then model utilization and collaboration improve, but proprietary information and internal details may be disclosed
Solution Approach 1:
The machine learning model is divided into multiple components: encrypted model weights, decryption keys, and inference code. The encryption system segments information access so that clients receive only encrypted model parameters without access to the actual model structure or training data, enabling model sharing while protecting proprietary information.
Solution Approach 2:
An encryption system acts as an intermediary between the model owner and the client. The system encrypts model weights using public key cryptography and manages key distribution, allowing the model to be shared and utilized without direct exposure of internal details to either party.
2Productivity
If trained machine learning models are shared with other organizations, then resource utilization improves, but the costs and proprietary nature of model development are compromised
Solution Approach 1:
Instead of sharing the actual trained model or its internal structure, the system creates encrypted copies of model weights that can be distributed and used for inference. The encryption ensures that while the model functionality is replicated and shared, the proprietary development details remain protected and cannot be extracted from the encrypted copies.
3Reliability
If machine learning models are encrypted to protect internal components, then security and privacy are improved, but model selection and utilization become more complex
Solution Approach 1:
The system performs preliminary encryption of model weights using public key cryptography before distribution. Decryption keys are pre-generated and distributed to authorized clients. This preliminary cryptographic setup enables secure model sharing without requiring complex real-time cryptographic operations during model selection and utilization phases.
Data Source
AI summary
Embodiments are directed towards classifying data. A machine learning (ML) engine may select an ML model that may employ a cryptographic multi-party computation (MPC) protocol based on model preferences, including a parameter model, provided by a client. A randomness engine may be employed to provide random values and other random values based on the MPC protocol such that the random values may be provided to the client and the other random values may be provided to an answer engine. Input values that correspond to fields in the parameter model may be provided by the client such that the input values may be based on the MPC protocol and the random values. The answer engine may be employed to provide partial results to the question based on the ML model, the input values, and the MPC protocol that may be provided to the client.