Joint Model Prediction With Linearized Activations for Secure MPC
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing secure multi-party computation (MPC) methods face challenges in efficiently and accurately performing multi-party joint model prediction while protecting private data, particularly due to the complexity of combining service prediction models with nonlinear activation functions.
Innovation Solution
Convert all nonlinear activation functions in the service prediction model into corresponding linear activation functions using Taylor expansion, ensuring a difference between the linear and nonlinear functions is less than a preset threshold, enabling secure multi-party computation for data exchange and prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If nonlinear activation functions are used in the service prediction model, then the model's predictive accuracy is improved, but the computational complexity and difficulty of implementing secure multi-party computation increase
Solution Approach 1:
The patent divides the domain of the nonlinear activation function into multiple segments, and for each segment, constructs a linear function through Taylor expansion that approximates the nonlinear function within that segment. This segmentation approach allows the complex nonlinear function to be replaced by simpler linear functions in different regions, reducing computational complexity while maintaining prediction accuracy through the piecewise linear approximation.
Solution Approach 2:
The patent changes the functional form parameter from nonlinear to linear by using Taylor expansion to approximate the nonlinear activation function. By expanding the nonlinear function around a point and keeping only the linear terms (first-order approximation), the complex nonlinear computation is transformed into simpler linear computation, which is more suitable for secure multi-party computation protocols.
2Reliability
If secure multi-party computation is implemented with nonlinear activation functions, then data privacy is protected, but computation efficiency and processing speed decrease
Solution Approach 1:
The patent substitutes the mechanical computation of nonlinear functions with linear function computations. Since linear functions have simpler mathematical properties and can be computed more efficiently using standard secure multi-party computation protocols (such as secret sharing and homomorphic encryption), this substitution significantly improves computation efficiency while maintaining the same level of data privacy protection.
3Productivity
If linear approximation is used to replace nonlinear activation functions, then computational efficiency is improved, but prediction accuracy may deteriorate
Solution Approach 1:
By dividing the domain into multiple segments and applying Taylor expansion in each segment, the patent achieves a better balance between accuracy and efficiency. The piecewise linear approximation captures the local behavior of the nonlinear function more accurately than a single global linear approximation, thereby maintaining prediction accuracy while improving computational efficiency.
Solution Approach 2:
The patent applies different linear approximations (derived from Taylor expansion at different points) to different segments of the domain. This local quality approach ensures that each segment has its own optimized linear function that best approximates the nonlinear function in that specific region, thereby maintaining high prediction accuracy across the entire domain while enabling efficient computation.
Data Source
AI summary
Embodiments of this specification provide privacy-preserving multi-party joint model prediction methods, systems, and apparatuses. A computing device converts all nonlinear activation functions in a service prediction model into corresponding linear activation functions, where any nonlinear activation function is converted into linear functions of several segments, any linear function is obtained by performing Taylor expansion on the nonlinear activation function at a Taylor expansion point in a corresponding segment, and in the segment, a difference between the linear function and the nonlinear activation function is less than a preset threshold. Then, a first participant and a second participant perform data exchange based on secure multi-party computation according to a logical operation indicated by the linear activation function included in the service prediction model, to jointly perform a prediction process of the service prediction model, where the secure multi-party computation is more applicable to calculation of the linear activation function.