Secure CNN Offloading from TEE to GPU via Intermediary Encryption
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional systems face challenges in securely and efficiently executing convolutional neural network (CNN) computations, particularly during training, due to limitations in offloading matrix multiplication from Trusted Execution Environments (TEEs) to GPUs, which leads to security exposures and slow performance.
Innovation Solution
A computerized framework that securely executes and offloads CNN computations within a cloud infrastructure by novel mechanisms for matrix multiplication from TEEs to GPUs, hiding and protecting weights, and eliminating pre-computation, thereby enhancing security and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If CNN computations are offloaded from TEE to GPU for efficient matrix multiplication, then computational speed is improved, but security is worsened due to exposure of sensitive data in untrusted environment
Solution Approach 1:
The patent introduces an intermediary encryption layer that mediates between the TEE and GPU. Sensitive data is encrypted before being transferred to the GPU, and the encryption keys remain securely stored in the TEE. This intermediary mechanism allows the GPU to perform computations on encrypted data without direct access to the plaintext, thus maintaining security while achieving computational speedup.
Solution Approach 2:
The system segments the computation process into two distinct phases: secure key management operations performed within the TEE, and computational operations performed on encrypted data on the GPU. This segmentation allows each component to operate in its optimal environment - the TEE for security-critical operations and the GPU for high-speed computation - while maintaining the security boundary.
2Productivity
If pre-computation is performed to optimize CNN operations, then computational efficiency is improved, but time and resource overhead is worsened
Solution Approach 1:
The patent performs preliminary actions of encrypting the data and preparing the computational framework before the actual CNN computations begin. The encryption schemes and secure communication channels are established in advance, allowing the main computational workload to proceed without repeated security overhead during the computation phases.
Data Source
AI summary
Disclosed are systems and methods for a computerized framework that provides an improved, secure computational environment between trusted and untrusted devices (e.g., a Trusted Execution Environment (TEE) and graphics processing unit (GPU), respectively) for executing and offloading Convolutional Neural Network (CNN) computations and operations. The disclosed framework can operate to perform such secure offloading and processing not only during inference computations of the CNN, but also during training of the CNN. The disclosed framework operates to enable training and execution of CNN models, whereby the data used for such operations are held securely while they are in use, in transit (e.g., between the TEE and GPU) and while in storage.