Homomorphic Encryption Processing Unit for Neural Network Operations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current homomorphic encryption techniques face limitations in processing encrypted data for neural networks, particularly with activation functions like sigmoid and ReLU, leading to high processing loads and communication volumes, and additive homomorphic encryption restricts calculations to integer multiplication, limiting process execution.
Innovation Solution
A processing apparatus and method that utilizes additive homomorphic encryption by introducing an adjustment multiplication value, allowing for integer processing of encrypted data, enabling wider processing capabilities by converting floating-point multipliers to fixed-point format and performing calculations in a ciphertext space, thereby expanding the range of executable processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If general homomorphic encryption is used to handle activation functions like sigmoid and ReLU, then the processing capability is improved, but the processing load becomes large and data volume increases
Solution Approach 1:
The patent transforms the processing approach by changing the parameter representation. Instead of using general homomorphic encryption that handles floating-point numbers directly, the invention converts all operations to integer arithmetic by scaling parameters. The activation functions and multipliers are represented as integers through fixed-point arithmetic, enabling the use of efficient additive homomorphic encryption while maintaining processing capability.
2Productivity
If additive homomorphic encryption is used to reduce processing load, then the processing load is reduced, but the executable process is limited to integer multiplication
Solution Approach 1:
The patent fundamentally changes the parameter representation from floating-point to integer-based fixed-point arithmetic. By scaling all numerical parameters (multipliers, activation function values) to integers, the system can use additive homomorphic encryption which is computationally efficient while still representing the full range of neural network operations. This parameter transformation eliminates the limitation to only integer multiplication.
Solution Approach 2:
The patent introduces an intermediary scaling factor that mediates between the requirements of additive homomorphic encryption (integer operations) and neural network processing (floating-point operations). The scaling factor acts as a converter, allowing floating-point multipliers to be represented as integer ratios, thereby enabling complex neural network operations through simple integer-based homomorphic encryption.
3Productivity
If data is transmitted from client to server for processing, then cloud resources are utilized, but information leakage risk increases
Solution Approach 1:
The patent applies preliminary encryption to the data before transmission. By encrypting the input data using additive homomorphic encryption before sending it to the server, the client ensures that the data remains confidential during transmission and processing. The server can perform computations on the encrypted data without being able to decrypt or view the actual information, thus preventing information leakage while still utilizing cloud computational resources.
Data Source
AI summary
A processing apparatus includes at least one processor configured to function as: an input unit that receives encrypted data based on homomorphic encryption as an input; and a process execution unit that executes a predetermined process by using the encrypted data while maintaining a secret state by encryption and includes one or more processing units. At least one of the processing units is a multiplication corresponding processing unit for executing a calculation in a ciphertext space corresponding to a processing of multiplying plaintext data by a predetermined multiplier. The multiplication corresponding processing unit executes a calculation in the ciphertext space corresponding to a calculation of multiplying the plaintext data by an adjustment multiplication value on first encrypted data input from a preceding stage and outputs resulting data. The adjustment multiplication value is obtained by multiplying the predetermined multiplier by a predetermined adjustment value for the multiplication corresponding processing unit and by performing an integer processing.