Neural Network Partition Hash Verification for Weight Integrity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In autonomous driving systems, there is a concern for ensuring the integrity and security of artificial neural networks (ANNs) to prevent errors and security hacks that could lead to safety issues, as existing methods do not effectively verify the integrity of synaptic weight and bias information.
Innovation Solution
The implementation of cryptography-based methods for ANN integrity verification, where cryptographic codes are generated and compared to ensure the accuracy of synaptic weights and biases before executing critical functions, issuing safety warnings for any integrity failures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If cryptography-based integrity verification is implemented for ANN synaptic weights and biases, then system safety and reliability are improved, but device complexity and computational overhead increase
Solution Approach 1:
The patent generates cryptographic codes (hashes) of the original ANN partition data and stores them in a verification database before the data is potentially tampered with or corrupted. This preliminary action enables later verification by comparing current data against the pre-stored cryptographic codes, ensuring integrity without adding complexity to the core ANN operation.
Solution Approach 2:
The patent introduces cryptographic codes as an intermediary layer between the ANN data and the verification process. Instead of directly verifying data integrity through complex comparisons, the system uses cryptographic hash functions to transform data into compact codes that serve as reliable intermediaries for integrity checking, simplifying the verification mechanism while maintaining high reliability.
2Reliability
If cryptographic verification codes are generated and stored for each ANN partition, then data integrity is improved, but memory usage and storage requirements increase
Solution Approach 1:
The patent extracts only the essential integrity verification information (cryptographic codes) from the complete ANN data and stores it separately in a verification database. This allows the main ANN operations to proceed with the original data while integrity verification uses only the extracted cryptographic codes, minimizing the impact on memory usage.
Solution Approach 2:
The patent transforms the original ANN partition data into cryptographic hash codes, fundamentally changing the data representation from complex neural network parameters to compact verification codes. This parameter transformation dramatically reduces the storage requirements for verification data while maintaining integrity verification capability.
3Reliability
If integrity verification operations are performed before executing critical functions, then safety is improved, but processing time and system performance are reduced
Solution Approach 1:
The patent performs integrity verification as a preliminary check before critical ANN functions are executed. By verifying data integrity upfront using pre-generated cryptographic codes, the system prevents potential safety issues from propagating through the system, ensuring that time-critical operations only process verified safe data.
Solution Approach 2:
The patent implements verification selectively for critical ANN partitions and functions rather than verifying all data at all times. This partial verification approach focuses computational resources on safety-critical operations, reducing overall processing time while maintaining safety where it matters most.
Data Source
AI summary
An example method comprises receiving a number of inputs to a system employing an artificial neural network (ANN), wherein the ANN comprises a number of ANN partitions each having respective weight matrix data and bias data corresponding thereto stored in a memory. The method includes: determining an ANN partition to which the number of inputs correspond, reading, from the memory the weight matrix data and bias data corresponding to the determined ANN partition, and a first cryptographic code corresponding to the determined ANN partition; generating, using the weight matrix data and bias data read from the memory, a second cryptographic code corresponding to the determined ANN partition; determining whether the first cryptographic code and the second cryptographic code match; and responsive to determining a mismatch between the first cryptographic code and the second cryptographic code, issuing an indication of the mismatch to a controller of the system.