Neural Network Model Integrity via Segmented Hashing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The large size of neural network models used in autonomous driving vehicles makes full image verification difficult during operation and pre-load verification impractical, especially since external memories lack hardware protection against unauthorized access.
Innovation Solution
The method involves segmenting model partitions into pre-configured memory sizes, hashing, and concatenating these segments offline, then storing them with the neural network models as an executable loadable file. This enables rapid validation during runtime by decrypting and comparing hash segments within the neural network's internal memory.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If full image verification is performed on neural network models during operation, then model integrity and security are ensured, but processing time and computational resources are excessively consumed due to the large model size
Solution Approach 1:
The patent divides the large neural network model into multiple smaller partitions, each with its own hash value stored in internal memory. During verification, only the relevant partition hashes are retrieved and compared against the loaded model data, rather than verifying the entire model at once. This segmentation approach maintains comprehensive integrity checking while reducing the time and computational resources required for verification.
2Reliability
If pre-load verification is performed on neural network models before operation, then model security is verified, but it is impractical due to the large model size and external memory limitations
Solution Approach 1:
The patent extracts the hash values of model partitions from the external memory and stores them in the internal memory of the neural network processor before operation begins. This extraction allows the verification process to work with small, manageable hash data rather than the full large-scale model, making pre-load verification practical and feasible while maintaining model security.
Solution Approach 2:
The patent performs the hashing of model partitions and storage of hash values in internal memory as a preliminary action before the actual model loading and verification during operation. This advance preparation ensures that when the model needs to be verified, the reference hash values are already available in fast internal memory, eliminating the need for time-consuming verification operations during critical runtime.
3Quantity of substance
If external memories are used to store large neural network models, then model storage capacity is sufficient, but hardware protection against unauthorized access is lacking
Solution Approach 1:
The patent introduces hash values as an intermediary layer between the external memory storing the model and the internal memory processing data. Instead of directly protecting the large model data in external memory (which lacks hardware protection), the system computes hash values from the model partitions and stores these compact representations in secure internal memory. This intermediary approach allows security verification without requiring hardware protection on the external storage itself.
Data Source
AI summary
A method and system is disclosed for protecting neural network models by segmenting partitions of the models into segments of pre-configured memory size, hashing the segmented models, and concatenating the hash segments. The concatenated hash segment may be further hashed, encrypted, and stored with the neural network models as an executable loadable file (ELF) in memories external to the neural network prior to the use of the models by the neural network. The models may include model weights of the inference layers and metadata. The model weights and the metadata may be hashed as separate hash segments and concatenated. Segmenting the models into segments of pre-configured memory size and hashing the segmented models offline prior to the operation of the neural network enables rapid validation of the models when the models are used in the inference layers during online operation of the neural network.