ML Notebook Cell Obfuscation via Hash-Based Masking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Federated learning environments face challenges in providing sufficient data privacy guarantees, leading to poor predictive performance due to the need for noise addition, and lack of fine-grained access control for on-premises assets when collaborating over the cloud.
Innovation Solution
A secure on-premises to cloud connector framework that uses machine learning notebooks with masked cells, encryption, and hash-based obfuscation to securely share models, datasets, and libraries, allowing controlled access and collaboration without exposing proprietary data to the cloud.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If federated learning adds noise to ensure data privacy, then data privacy is improved, but predictive model performance deteriorates
Solution Approach 1:
The patent introduces an on-premises gateway as an intermediary component that mediates between the federated learning system and on-premises assets. This gateway enables secure access to local data and models without requiring noise addition, as the actual data remains protected behind the gateway while still allowing legitimate federated learning operations to proceed with high accuracy
Solution Approach 2:
The system segments access control at the cell level within notebook files, allowing fine-grained permissions for different users and groups. This segmentation enables selective sharing of specific code cells, data assets, or model components without exposing entire datasets, thereby maintaining privacy without degrading model performance through noise
2Adaptability or versatility
If cloud-based federated learning is implemented, then collaboration capability is improved, but data security and access control deteriorate
Solution Approach 1:
The on-premises gateway acts as a trusted intermediary that enables cloud-based collaboration while maintaining data security. It allows federated learning participants to access and train models on local data without directly exposing the data to the cloud, thus preserving security while enabling collaboration
Solution Approach 2:
The system implements local quality control by allowing different security and access policies for different on-premises assets. Each notebook file, code cell, data asset, and model can have customized permission settings, enabling fine-grained control over what is shared collaboratively and what remains private
3Adaptability or versatility
If on-premises assets are shared over cloud, then collaboration is improved, but fine-grained access control is lost
Solution Approach 1:
The patent implements local quality by applying different access control policies to different levels of the notebook structure. Each notebook file, code cell, data asset, and model can have independent permission settings, enabling fine-grained control where specific cells or assets can be shared with specific users or groups while others remain private
Solution Approach 2:
The system segments the notebook file into discrete cells, each of which can be independently controlled. This segmentation allows owners to grant access to individual cells or groups of cells rather than forcing all-or-nothing sharing, thereby enabling precise access control that supports collaboration without exposing unnecessary assets
Data Source
AI summary
Embodiments securely share a machine learning (“ML”) notebook, comprising a plurality of cells, over a cloud network. Embodiments receive the ML notebook with one or more of the cells designated as a masked cell. Embodiments encrypt the masked cells and hash the masked cell using a corresponding hash. Embodiments store the hashed masked cell with a corresponding one or more identities of users who can use the hash to execute the masked cell.