Encrypted DNN Inference Using Quantization and Integer FHE
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current FHE-based DNN models face significant computational overhead and inefficiencies, particularly with non-linear operations like ReLU and Sigmoid activation functions, making them impractical for real-world applications.
Innovation Solution
A method involving encoding and decoding of messages using quantization parameters, noise-aware fine-tuning, and a modified FHEW-variant DNN model with integer-based computations and functional bootstrapping to enhance efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Fully Homomorphic Encryption is integrated with Deep Neural Networks, then data security and privacy are improved, but computational overhead and computation time increase significantly
Solution Approach 1:
The patent applies parameter changes by converting the DNN model from floating-point to integer format, adjusting precision parameters, and modifying quantization settings to optimize the balance between security requirements and computational efficiency in the FHE environment
Solution Approach 2:
The patent performs preliminary actions by pre-processing the DNN model before encryption, including model quantization, parameter conversion to integer format, and optimization of computational graphs to minimize FHE overhead during actual inference
2Reliability
If FHE is used for non-linear operations in DNNs, then semantic security is improved, but computational cost and complexity increase
Solution Approach 1:
The patent changes parameters by converting activation functions to integer-compatible forms, adjusting precision levels, and modifying computational parameters to reduce the complexity of evaluating non-linear operations under FHE while maintaining security guarantees
Solution Approach 2:
The patent substitutes complex floating-point non-linear operations with simplified integer-based approximations and lookup tables that are more efficient in the FHE context, replacing computationally expensive mechanical operations with more suitable integer arithmetic
3Reliability
If FHE-based DNN computation is implemented, then data privacy is protected, but productivity and computational efficiency deteriorate
Solution Approach 1:
The patent changes computational parameters including precision levels, quantization granularity, and integer data types to optimize the trade-off between privacy protection through FHE and computational efficiency for specific application scenarios
Solution Approach 2:
The patent applies partial action by selectively applying FHE only to critical computational paths and data structures where privacy is most important, rather than encrypting the entire model and all operations uniformly, thus improving overall efficiency
Data Source
AI summary
An inference method for encrypted deep neural network model is executed by a computing device and includes: encoding a message according to a quantization parameter to generate a plaintext, encrypting the plaintext according to a private key to generate a ciphertext, sending the ciphertext to a deep neural network model to generate a ciphertext result, decrypting the ciphertext result according to the private key to generate a plaintext result, and decoding the plaintext result according to the quantization parameter to generate an inference result.