Neural Network Weight Encryption and On-Chip Decryption
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network processors lack effective mechanisms for protecting weight security and intellectual property, leading to issues such as weight value leakage and flood propagation during neural network processing.
Innovation Solution
A weight management method and system that encrypts trained weight data using an AES encryption algorithm, decrypts it on-chip in real-time for neural network operations, and stores the decrypted data securely, ensuring the integrity and security of the original data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If weight data is stored in plaintext for neural network processing, then processing speed and efficiency are improved, but weight security and intellectual property protection deteriorate
Solution Approach 1:
The patent applies preliminary action by encrypting weight data before it is loaded into the neural network processor. The encryption is performed in advance during the weight preparation phase, so that when the weights are loaded for processing, they are already in encrypted form. This eliminates the need to switch between plaintext and encrypted states during processing, maintaining both security and processing efficiency.
Solution Approach 2:
The patent introduces an intermediary decryption unit that acts as a mediator between the encrypted weight storage and the neural network processing units. This decryption unit converts encrypted weights to plaintext only when needed for specific operations, and immediately re-encrypts or secures them afterward. This intermediary layer protects the weight data while enabling necessary processing operations.
2Reliability
If weight data is encrypted to protect security, then weight leakage prevention is improved, but processing complexity and computational overhead increase
Solution Approach 1:
The patent merges the decryption functionality directly into the neural network processor architecture. Instead of having separate decryption hardware or software layers that add complexity, the decryption unit is integrated within the processor itself, allowing encrypted weights to be decrypted and processed in a unified, streamlined manner without requiring complex external systems.
Solution Approach 2:
The neural network processor performs self-service by including its own decryption capability within its architecture. The processor can independently decrypt its own weight data without requiring external decryption systems, reducing overall system complexity and eliminating the need for complex interactions between separate security and processing subsystems.
3Measurement precision
If weight data is decrypted for processing operations, then neural network operation accuracy is improved, but exposure to weight leakage risks increases
Solution Approach 1:
The patent applies local quality by decrypting weight data only in specific localized contexts where it is absolutely necessary for processing operations. Rather than maintaining all weights in plaintext or encrypting them uniformly, the system selectively decrypts only the specific weight values needed for current computations, keeping other weights encrypted. This localized decryption minimizes exposure time and scope, reducing leakage risk while maintaining processing accuracy.
Solution Approach 2:
The patent implements skipping by rapidly decrypting weight data only when absolutely necessary for a computation step, performing the required operation immediately, and then transitioning back to encrypted state. This 'rush through' approach minimizes the time window during which decrypted weights are vulnerable to leakage, allowing the system to quickly pass through the vulnerable decrypted state without lingering in it.
Data Source
AI summary
The present invention relates to a weight management method and system for neural network processing. The method includes two stages, i.e., off-chip encryption stage and on-chip decryption stage: encrypting trained neural network weight data in advance, inputting the encrypted weight into a neural network processor chip, and decrypting the weight in real time by a decryption unit inside the neural network processor chip to perform related neural network calculation. The method and system realizes the protection of weight data without affecting the normal operation of a neural network processor.