Risk Evaluation Apparatus Using Loss Function Regression for Adversarial Example Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for creating Adversarial Examples require knowledge of the machine learning model's loss function, which is often not easily accessible in real-world scenarios, making it difficult to evaluate the risk of a machine learning model without performing a model extraction attack.
Innovation Solution
A risk evaluation apparatus that records predetermined loss functions and data-label pairs, uses nonparametric regression to determine a regression model of the loss function in the vicinity of the data, and creates Adversarial Examples using this regression model to evaluate the risk without requiring knowledge of the model's internal mechanisms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If model extraction attack is performed to create Adversarial Examples, then the ability to create effective attacks is improved, but the complexity of the attack process and requirement for internal model knowledge increases
Solution Approach 1:
The patent introduces a loss function regression model as an intermediary that approximates the true loss function without requiring direct access to it. This intermediary model enables adversarial example generation while avoiding the need for model extraction attacks, thus reducing attack complexity while maintaining reliability
Solution Approach 2:
The patent replaces the mechanical process of model extraction attack with a regression-based approximation approach. Instead of directly extracting model information through complex attacks, the system uses nonparametric regression to estimate the loss function behavior, substituting a simpler statistical approach for a complex adversarial process
2Measurement precision
If model extraction attack is required to create Adversarial Examples, then accurate risk evaluation is improved, but the ease of operation deteriorates due to need for internal model knowledge
Solution Approach 1:
The system performs self-service by using only the machine learning model's input-output behavior to generate adversarial examples. The loss function regression model is trained automatically using publicly accessible predictions, eliminating the need for operators to have internal model knowledge while maintaining evaluation accuracy
Solution Approach 2:
The patent performs preliminary action by pre-training the loss function regression model using the machine learning model's predictions on various inputs. This preliminary training phase captures the loss function's characteristics, enabling subsequent adversarial example generation without requiring direct access to the original loss function or model internals
Data Source
AI summary
The risk evaluation apparatus evaluates the risk of a machine learning model. The risk evaluation apparatus includes a recording unit, a loss function regression model acquirer, an attack noise addition unit, an error acquisition unit, and an evaluation unit. The recording unit records a set of predetermined loss functions and a set of pairs of data and labels predetermined. The loss function regression model acquirer determines a regression model of the loss function in the vicinity of data by nonparametric regression. The attack noise addition unit creates attack data that is an Adversarial Example using the regression model. The error acquisition unit determines the error between the output of the machine learning model when the data is input and the output of the machine learning model when the attack data is input. The evaluation unit evaluates the risk based on a set of errors.