Neural Cryptosystem Secret Sharing via DNN Layer Partitioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural cryptography methods require user-provided keys and are inefficient in protecting deep neural network (DNN) models, as existing runtime protection methods consume excessive resources or have limited bandwidth, and cannot handle non-polynomial operations effectively.
Innovation Solution
Partitioning the DNN model into public and private layers, where public layers act as a public key and private layers as a private key, using adversarial training to ensure that only access to both layers allows full model functionality, and deploying the private layers in a protected execution environment to prevent unauthorized access.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If user-provided keys are used in neural cryptography, then the system can implement encryption/decryption, but the system requires additional user input and cannot achieve seamless integration with existing DNNs
Solution Approach 1:
The DNN model generates its own cryptographic keys (public and private layers) autonomously during the key generation process, eliminating the need for user-provided keys. The model serves itself by creating the cryptographic components it needs for secure inference.
2Reliability
If runtime protection methods are used to protect DNN models, then security is improved, but resource consumption increases excessively
Solution Approach 1:
The system performs protective actions in advance by partitioning the DNN model into public and private layers and securing the private layers before deployment. The private layers are prepared and protected beforehand, so that during runtime, the model can operate securely without consuming excessive resources for protection.
3Reliability
If existing runtime protection methods are used, then some security is provided, but bandwidth is limited and operations cannot be handled effectively
Solution Approach 1:
The DNN model is segmented into public layers and private layers, where public layers handle general operations and private layers contain protected functionality. This segmentation allows the system to provide security while maintaining full operational capability, as the private layers can process any operation that requires protection without bandwidth limitations.
4Ease of operation
If the DNN model is made fully accessible, then complete functionality is available, but unauthorized access and model compromise occur
Solution Approach 1:
Different parts of the DNN model have different security properties: public layers are fully accessible for general operations, while private layers have restricted access for protected operations. This local differentiation of access rights allows the model to maintain both accessibility and security, with each layer having the appropriate level of protection for its function.
Data Source
AI summary
Partitioning a deep neural network (DNN) model into one or more sets of one or more private layers and one or more sets of one or more public layers, a set of one or more private layers being at least one key in a cryptographic system; and deploying the partitioned DNN model on one or more computing systems.