Secure Deep Learning via Data Shuffling and Offsetting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional Secure Multi-Party Computation (SMPC) techniques require high precision circuits to operate on ciphertexts generated using Homomorphic Encryption, which are not available in typical Deep Learning Accelerators (DLAs), limiting their ability to perform outsourced deep learning computations.
Innovation Solution
The technique involves shuffling randomized data parts across multiple samples, allowing external parties to perform computations without reconstructing the original data, and using an offset key for enhanced privacy protection, enabling DLAs without high precision circuits to participate in computations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Homomorphic Encryption is used for Secure Multi-Party Computation, then data privacy is protected, but high precision circuits are required which are not available in typical Deep Learning Accelerators
Solution Approach 1:
The patent segments data samples into multiple parts and distributes them to different external parties. Each party receives only a portion of the data, making it impossible to reconstruct the original data without all parts. This segmentation approach enables secure computation without requiring high precision circuits for encryption, as the security is achieved through data distribution rather than cryptographic complexity.
Solution Approach 2:
The patent introduces an intermediary shuffling mechanism that randomizes the distribution of data parts among external parties. The shuffling acts as a mediator that prevents any single party from obtaining meaningful information while still allowing computation to proceed. This intermediary layer enables privacy protection through randomized distribution rather than through encryption circuits.
2Productivity
If data is outsourced for computation, then computing efficiency is improved, but data security risks increase
Solution Approach 1:
By dividing data into multiple segments and distributing them to different external computing parties, the patent enables efficient outsourced computation while ensuring that no single party has access to the complete data set. This segmentation allows computation to be performed on distributed data parts without exposing the full data, thus maintaining security while achieving computing efficiency.
Solution Approach 2:
The patent applies different security measures to different parts of the data. Each data part is protected through its specific distribution and shuffling pattern, allowing external parties to perform computations on their local parts without needing access to other parts. This local quality approach enables secure outsourced computation by making each local data portion inherently protected.
3Reliability
If encryption keys are lengthened for stronger security, then brute-force attack resistance is improved, but circuit precision requirements increase beyond DLA capabilities
Solution Approach 1:
Instead of using long encryption keys that would require high precision circuits, the patent segments data into multiple parts and uses distribution-based security. The security against brute-force attacks is achieved through the combinatorial complexity of reconstructing data from segmented parts rather than through key length, thus avoiding the need for high precision circuits in Deep Learning Accelerators.
Data Source
AI summary
Protection of access to data samples in outsourcing deep learning computations via shuffling parts. For example, each data sample can be configured as the sum of a plurality of randomized parts. At least some of the randomized parts can be applied an offset operation to generate modified parts for outsourcing. Such parts from different data samples are shuffled and outsourced to one or more external entities to apply a deep learning computation. The deep learning computation is configured to allow change of the order between applying the summation and applying the deep learning computation. Thus, results of the external entities applying the deep learning computation to their received parts can be shuffled back for a data sample to apply reverse offset and summation. The summation provides the result of applying the deep learning computation to the data sample.