Non-uniform Tensor Splitting for Secure DLA Outsourcing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional Secure Multi-Party Computation (SMPC) techniques using Homomorphic Encryption are not compatible with Deep Learning Accelerators (DLAs) due to the requirement for high precision circuits, which are not typically available in these devices, limiting their ability to perform operations like multiplication and accumulation.
Innovation Solution
The proposed solution involves shuffling and partitioning data samples and model parts into randomized parts, allowing external entities with DLAs to perform computations without the need for high precision circuits, using techniques such as offsetting and Homomorphic Encryption to enhance data privacy and protect against brute-force attacks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Homomorphic Encryption is used for Secure Multi-Party Computation, then data privacy is protected, but compatibility with Deep Learning Accelerators is lost due to high precision circuit requirements
Solution Approach 1:
The patent divides the tensor into multiple partitions along different dimensions (rows, columns, depth). Each partition is processed separately by different external entities, allowing the computation to be distributed across multiple DLAs without requiring any single entity to hold the complete encrypted tensor, thus maintaining both security and DLA compatibility
Solution Approach 2:
Different partitioning strategies are applied to different portions of the tensor based on their security requirements and computational characteristics. The patent applies selective partitioning where only certain portions requiring security are divided, while other portions can be processed more efficiently, optimizing both security and performance
2Reliability
If data is split into multiple partitions for secure computation, then security against brute-force attacks is enhanced, but computation workload increases
Solution Approach 1:
The patent combines multiple partitioning dimensions (row partitioning, column partitioning, depth partitioning) to create a multi-layered security approach. By partitioning along multiple dimensions simultaneously, the security is exponentially enhanced while the computational overhead is distributed and managed efficiently across external entities
Solution Approach 2:
The patent applies partitioning selectively to only those portions of the tensor that require security protection, rather than uniformly partitioning the entire tensor. This partial action approach reduces unnecessary computation overhead while maintaining security for critical data portions
3Ease of operation
If uniform partitioning is used for tensor splitting, then simplicity is maintained, but security against reconstruction attacks is reduced
Solution Approach 1:
The patent employs asymmetric partitioning where different external entities receive partitions of different sizes and characteristics. Some entities receive larger partitions while others receive smaller ones, and the partition boundaries are irregular rather than uniform. This asymmetry prevents any single entity from easily reconstructing the original tensor and complicates attack vectors
Data Source
AI summary
Protection of access to values of elements in a tensor in outsourcing deep learning computations. For example, the tensor in the computation of an artificial neural network can be partitioned into portions. Some of the portions can be selected for splitting into parts, such that the sum of a set of parts is equal to a respective portion being split to generate computing tasks. Each computing task is configured to operate based on a portion of the tensor or a part of a portion of the tensor. Some of the portions may share common parts. The computing tasks can be generated according to unique parts to eliminate duplicative computing efforts. The computing tasks can be shuffled for distribution out of order to external entities. The result to operate on the tensor can be obtained from results, received back from the external entities, of the outsourced computing tasks.