Homomorphic Encryption Floating-Point AI Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fully homomorphic encryption methods, such as CKKS and FHEW, face limitations in supporting floating-point operations and achieving accurate neural network computations due to approximations and integer encryption schemes.
Innovation Solution
A privacy computing method based on homomorphic encryption that uses a floating-point homomorphic encryption data structure with a ciphertext mantissa, exponent parameter, and gain parameter to enable accurate operations on encrypted data within artificial intelligence models, such as neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If CKKS approximate computing homomorphic encryption algorithm is used to support nonlinear function calculation, then floating-point operations are enabled, but a large number of multiplication operations are required and accuracy is reduced due to Taylor expansion approximation
Solution Approach 1:
The patent transforms the encryption scheme parameters by converting integer-based homomorphic encryption into a floating-point system using mantissa-exponent-gain representation. This parameter transformation enables accurate floating-point arithmetic operations while maintaining homomorphic properties, resolving the contradiction between supporting floating-point operations and maintaining calculation accuracy.
2Reliability
If FHEW fully homomorphic encryption based on GSW technology is used, then integer encryption is achieved, but floating-point operations are not supported and neural network operations cannot be effectively performed
Solution Approach 1:
The patent fundamentally changes the encryption parameter structure by introducing a floating-point homomorphic encryption scheme that represents numbers as combinations of mantissa, exponent, and gain parameters. This allows the system to perform neural network operations on encrypted floating-point data while maintaining the security properties of homomorphic encryption, thus resolving the contradiction between encryption reliability and adaptability to floating-point operations.
3Productivity
If data is transmitted to model owner for calculation, then computation can be performed, but data owner faces privacy leakage risk
Solution Approach 1:
The patent introduces homomorphic encryption as an intermediary mechanism that allows computations to be performed on encrypted data. The encryption scheme acts as a mediator between the data owner and model owner, enabling the model owner to process data without accessing the plaintext, thus maintaining privacy while enabling computation.
4Ease of operation
If model is transmitted to data owner for calculation, then data owner can perform computation, but model owner faces privacy leakage risk
Solution Approach 1:
The patent uses homomorphic encryption as an intermediary that protects the model owner's intellectual property. By requiring the data owner to use encrypted data with the model, the system ensures that the model owner's proprietary model parameters remain confidential while still allowing the data owner to perform computations locally.
Data Source
AI summary
The present disclosure provides a privacy computing method based on homomorphic encryption, which includes steps as follows. The ciphertext data is received, where the ciphertext data has a floating-point homomorphic encryption data structure, and the floating-point homomorphic encryption data structure of the ciphertext data includes the ciphertext mantissa, exponent parameter and gain parameter. The gain parameter sets the precision of the floating point corresponding to the ciphertext mantissa. The exponent parameter is adapted to multiplication or division. The artificial intelligence model performs operations on the ciphertext data to return the ciphertext result.