Trust Environment AI Device for Secure DNN Execution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI neural network frameworks in mobile and embedded devices face security challenges due to limited memory protection and high data encryption/decryption overhead, which slow down DNN execution and increase memory access times.
Innovation Solution
An AI device based on a trust environment that encrypts data in a trust space and reduces memory accesses by performing direct convolution-based neural network computations, utilizing cryptographic hardware for offloading and intra-layer pipelining to overlap neural network computations with encryption and decryption operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data is encrypted in secure on-chip memory to protect from physical attacks, then security is improved, but memory access speed deteriorates and DNN execution time increases
Solution Approach 1:
The patent divides memory into secure on-chip memory for encrypted data storage and separate processing units. The cryptographic processing front-end processor handles encryption/decryption operations independently from the main neural network processor, allowing secure memory access without blocking the main computation pipeline.
Solution Approach 2:
The cryptographic processing front-end processor acts as an intermediary between the secure encrypted memory and the neural network processor. It decrypts input data before processing and encrypts output data after processing, mediating the interaction between secure storage and computation while maintaining security boundaries.
2Reliability
If data encryption and decryption operations are performed on the processor, then security is improved, but processor resources are consumed and DNN execution time increases
Solution Approach 1:
The patent extracts the cryptographic processing functions from the main neural network processor and implements them in a separate cryptographic processing front-end processor. This extraction removes the security-related encryption/decryption overhead from the main processor, allowing it to focus exclusively on neural network computations.
Solution Approach 2:
The cryptographic processing front-end processor handles its own encryption and decryption operations independently, serving the neural network processor with decrypted data without requiring the main processor to perform security operations. This self-service approach prevents security operations from consuming neural network processing resources.
3Reliability
If traditional memory access patterns are used for encrypted data, then security is maintained, but the number of memory accesses increases and execution time increases
Solution Approach 1:
The cryptographic processing front-end processor performs decryption of input data before it is needed by the neural network processor, and encryption of output data after processing is complete. This preliminary action ensures that the main processor receives decrypted data ready for immediate processing, eliminating the need for the processor to pause and perform decryption operations during the computation pipeline.
Data Source
AI summary
An artificial intelligence (AI) device based on a trust environment, includes a first type memory configured to transmit encrypted input data and receive encrypted output data, and a trust AI processing unit configured to operate in a trust space and perform AI computation of the encrypted input and output data. The trust AI processing unit includes: a cryptographic processing front-end processor configured to generate decrypted input data through decryption of the encrypted input data and perform encryption of non-encrypted output data to generate the encrypted output data, a second type memory configured to provide a buffer for the decrypted input data and the non-encrypted input data, and a processor configured to perform a neural network computation based on the decrypted input data to generate the non-encrypted output data.