FHE Network Conversion via Expansion Factors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current fully homomorphic encryption (FHE) schemes, particularly torus-based schemes like TFHE, face challenges in performing operations directly on encrypted values, especially with non-integral values, which are common in neural networks and other computations, leading to inefficiencies and the need for programmable bootstrapping.
Innovation Solution
The method involves converting real-valued matrix operations and other computations into integer-valued operations by using expansion factors and rounding, allowing direct performance of operations on encrypted values without programmable bootstrapping, while adjusting other operations to accommodate scaled inputs and outputs, and iteratively optimizing expansion factors for accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If programmable bootstrapping is used to perform operations on encrypted non-integral values, then the functionality and versatility of FHE computations are improved, but the computational complexity and execution time increase significantly
Solution Approach 1:
The patent transforms the parameter domain by converting non-integral values (real numbers) into integral values through scaling and rounding operations. This parameter change allows the system to operate in the integer domain where simpler FHE operations can be applied, avoiding the need for complex programmable bootstrapping while maintaining computational functionality.
Solution Approach 2:
The patent introduces intermediate scaling factors as mediators between the original non-integral values and the integer operations. These scaling factors serve as intermediaries that enable the conversion from real number domain to integer domain, allowing standard FHE operations to process what would otherwise require programmable bootstrapping.
2Measurement precision
If programmable bootstrapping is used to handle non-integral values in FHE, then the accuracy of computations is maintained, but the execution time and computational resources increase
Solution Approach 1:
The patent performs preliminary scaling and rounding operations on the input values before they enter the main FHE computation pipeline. By pre-processing the data to convert non-integral values into integral values with appropriate scaling, the system avoids the need for time-consuming programmable bootstrapping operations during the actual computation, thereby improving execution time while preserving accuracy.
Solution Approach 2:
The patent changes the numerical parameter domain from real numbers to integers through scaling and rounding. This parameter transformation enables the use of efficient integer-based FHE operations instead of programmable bootstrapping, significantly reducing execution time while maintaining computational accuracy through careful selection of scaling factors.
3Productivity
If real-valued operations are converted to integer-valued operations using expansion factors, then the efficiency of FHE computations is improved, but the manufacturing complexity of the conversion process increases
Solution Approach 1:
The patent segments the conversion process into distinct, manageable steps: scaling the real-valued inputs by expansion factors, rounding to obtain integer values, and then performing standard FHE operations. This segmentation of the conversion process makes it easier to implement and maintain, reducing the manufacturing complexity despite the efficiency gains.
Data Source
AI summary
Some embodiments are directed to a computer-implemented method for converting a first computation network of operations arranged to compute a function into a second computation network of FHE operations arranged to compute the same function. For example, a set of expansion factors (αi) may be determined for matrix operations in the first computation network. Real-valued matrices may be converted by scaling the real-valued matrices with their corresponding expansion factor and rounding. An accuracy measure for the second computation network can be iteratively optimized.