Secure SVM Kernel Evaluation via Segmented Exponential Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing secured data classification methods using support vector machines with kernels like Gaussian or hyperbolic tangent face challenges in securely calculating exponentials, leading to incomplete security and non-exact results due to the complexity of exponentiation operations in multiparty computation protocols.
Innovation Solution
A method involving a client and server unit performing cryptographic calculations by generating reversible data for multiplication, applying oblivious transfer protocols, and multiplying intermediate results to evaluate kernel functions securely, ensuring no party learns information about the other's data, particularly by breaking down functions into elementary exponential components dependent on single bits of the classified datum.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If secured calculation methods using Taylor expansion or random masking are applied to compute exponential functions in SVM kernels, then some level of privacy is preserved, but the security is incomplete and results become non-exact due to the complexity of exponentiation operations
Solution Approach 1:
The patent segments the exponential function computation into multiple elementary operations that can be performed securely. Instead of computing the entire exponential function at once, the method breaks it down into smaller sub-functions that are evaluated separately using secure protocols, then combined to produce the final result. This segmentation allows each sub-computation to maintain security while preserving overall accuracy.
Solution Approach 2:
The patent introduces intermediary values and protocols to facilitate secure computation. Random masking values are introduced as intermediaries to hide the actual input data during computation. Oblivious transfer protocols serve as intermediaries to enable parties to exchange information without revealing their private data. These intermediaries allow the exponential function to be computed securely while maintaining precision.
2Adaptability or versatility
If complex exponentiation operations are performed in multiparty secured computation protocols, then kernel functions like Gaussian and hyperbolic tangent can be evaluated, but the computational complexity increases significantly
Solution Approach 1:
The patent applies segmentation by dividing the complex exponential function into multiple simpler sub-functions. Each sub-function can be evaluated using standard secure computation protocols, avoiding the need for complex custom protocols for the entire exponential operation. This reduces the overall protocol complexity while maintaining the ability to evaluate Gaussian and hyperbolic tangent kernels.
Solution Approach 2:
The patent performs preliminary actions by pre-computing certain values and preparing secure evaluation tables before the actual classification task. This allows the exponential function to be evaluated more efficiently during the secure computation phase, reducing the complexity of the main protocol while still supporting various kernel functions.
3Loss of information
If random masking values are used to secure the computation of f(x,y)+r, then privacy is partially protected, but information leakage can still occur through inference from the masked results
Solution Approach 1:
The patent uses oblivious transfer protocols as intermediaries to securely exchange the masked values between parties. This ensures that the random masking values are properly protected and that the final result can be accurately reconstructed without information leakage. The intermediary protocol guarantees that neither party can infer the other's private data from the masked results.
Solution Approach 2:
The patent changes the parameters of the masking values to ensure they are sufficiently large and randomly distributed to prevent inference attacks. By adjusting the distribution and magnitude of the random masking values, the patent maintains both privacy protection and result accuracy, preventing parties from inferring sensitive information while preserving the correctness of the computation.
Data Source
AI summary
A method for calculating a function as a sub-function fX of a datum of a client, a sub-function fY of a datum, and a product of n indexed sub-functions fi of both data by randomly generating, n indexed invertible data ri from the set with m being a prime number, generating, for each i from 1 to n, a set for which each element is formed by a product of a datum ri with a possible result of the sub-function of two variables fi evaluated in both data, applying an oblivious transfer protocol between the client and a server so that the client recovers, for each i an intermediate datum ti equal to: ti=ri×fi(xi,Y), obtaining, by the client a result T from intermediate data such that: T=fX(X′)×Πi=1nti, obtaining, by the server a result R from inverted data such that: R=fY(Y)×Πi=1nri−1 using the results in a cryptographic application.