SecureHD Hyperdimensional Vector Encoding for Privacy-Preserving Cloud ML
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
IoT applications face challenges in securely offloading computation tasks to the cloud due to privacy and security concerns, with existing methods like homomorphic encryption and secure Multi-Party Computation being computationally expensive and inefficient.
Innovation Solution
A framework called SecureHD that uses high-dimensional vector encoding and decoding to securely perform machine learning tasks without decrypting data, allowing secure collaborative training and inference while reducing computational burden and maintaining data privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If homomorphic encryption is used to encrypt data before cloud processing, then data privacy is protected, but computation time increases significantly (14 days for MNIST dataset)
Solution Approach 1:
The patent extracts the encryption/decryption operations from the cloud processing workflow and relocates them to edge devices. Only encrypted data is transmitted to the cloud for processing, eliminating the need for cloud-based decryption and re-encryption operations that cause computational delays.
Solution Approach 2:
The system segments the machine learning workflow into encryption at the edge, secure processing in the cloud, and decryption at the edge. This segmentation allows the cloud to process only encrypted data without bearing the computational overhead of cryptographic operations.
2Reliability
If homomorphic encryption is applied to enable secure cloud processing, then data security is maintained, but data size increases by 28 times
Solution Approach 1:
The patent extracts the bulky encrypted data storage requirement from the cloud environment and places it at the edge devices where encryption and decryption occur. The cloud only handles processing of encrypted data, avoiding the need to store large volumes of encrypted data.
3Reliability
If secure Multi-Party Computation techniques are used for distributed learning, then data privacy is protected across multiple users, but communication overhead and computational cost increase significantly
Solution Approach 1:
The patent introduces encrypted data as an intermediary representation that enables secure collaborative learning without requiring complex multi-party computation protocols. Each user encrypts their data locally, and the cloud processes encrypted data from multiple users without needing to coordinate secure computations between them.
4Measurement precision
If machine learning models require unencrypted plaintext data for training, then model accuracy is maintained, but sensitive information is exposed to untrustworthy cloud systems
Solution Approach 1:
The patent uses encrypted data as a disposable intermediate representation that can be processed by the cloud without compromising security. The encrypted data serves its purpose during processing and is then discarded, never being decrypted in the untrustworthy cloud environment.
Data Source
AI summary
A computing system can include a plurality of clients located outside a cloud-based computing environment, where each of the clients may be configured to encode respective original data with a respective unique secret key to generate data hypervectors that encode the original data. A collaborative machine learning system can operate in the cloud-based computing environment and can be operatively coupled to the plurality of clients, where the collaborative machine learning system can be configured to operate on the data hypervectors that encode the original data to train a machine learning model operated by the collaborative machine learning system or to generate an inference from the machine learning model.