Deep Learning Neural Network Storage Security via Dummy Nodes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep-learning neural networks face security risks due to the potential for unauthorized access and inference when model parameters are not adequately encrypted, allowing hackers to utilize the network even with partial parameter access.
Innovation Solution
A storage and inference method is introduced that establishes dummy nodes in a deep-learning neural network, separating model parameters and dummy node parameters across different storage areas, and deleting or setting interconnections between dummy nodes to zero during inference, thereby maintaining network architecture integrity and enhancing encryption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If model parameters are stored in a single storage area, then storage and retrieval are simple, but security is compromised as hackers can use partial parameters for inference
Solution Approach 1:
The patent divides model parameters into two distinct segments: real node parameters and dummy node parameters. These segments are stored in separate storage areas (first storage area for real parameters, second storage area for dummy parameters). This segmentation prevents hackers from obtaining complete model information even if they access one storage area, thereby improving security while maintaining manageable storage structure
Solution Approach 2:
Dummy nodes are introduced as intermediary elements between the real model parameters and the inference process. These dummy nodes with zero weights act as mediators that obscure the real parameters during storage while being automatically excluded during inference, thus enhancing security without affecting the actual inference functionality
2Reliability
If dummy nodes are added to the neural network, then security and encryption are enhanced, but the network architecture becomes more complex
Solution Approach 1:
Dummy nodes are designed as disposable computational elements that serve their security function during the storage and retrieval phase, then are effectively discarded during inference by setting their weights to zero. They provide encryption/security benefits without requiring permanent structural changes to the neural network architecture
Solution Approach 2:
The patent employs parameter changes by setting dummy node weights to zero during inference operations. This parameter transformation allows the same neural network structure to function with enhanced security during storage while maintaining original functionality during inference, without requiring architectural redesign
3Stability of the object's composition
If interconnections between dummy nodes are deleted or set to zero, then the original network architecture is preserved, but additional processing steps are required before inference
Solution Approach 1:
The patent applies preliminary action by pre-setting the weights of dummy node interconnections to zero during the model deployment phase. This preliminary configuration ensures that when inference is performed, no additional complex processing is needed—the dummy nodes are already configured to be computationally inert, thus preserving architecture stability without significant productivity loss
Data Source
AI summary
A storage and inference method for a deep-learning neural network comprises steps: establishing dummy nodes in a first artificial neural network to form a second artificial neural network; storing model parameters of the second artificial neural network in a first storage area, and storing parameters of the dummy nodes in a second storage area; and in inference, respectively retrieving the model parameters of the second artificial neural network and the parameters of the dummy nodes from the first storage area and the second storage area simultaneously; deleting interconnections between the dummy nodes of the second artificial neural network or setting the interconnections between the dummy nodes of the second artificial neural network to 0 according to the parameters of the dummy nodes before inference. The present invention prevents ANN from be deciphered through respectively storing model parameters and parameters of the dummy nodes in different locations.