Robustness Measurement Device for ML Models Without Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for evaluating the noise resistance of machine learning models are limited, as they require correct answer data for inference and cannot assess models without training data, and they do not effectively handle object detection tasks.
Innovation Solution
A robustness measurement device that acquires input images, adds noise, generates noise-added images, and measures model robustness by grouping inference results based on overlap ratios, allowing for accurate noise resistance evaluation regardless of the presence of training data, and supports object detection processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If correct answer data is used for evaluation, then noise resistance can be measured accurately, but evaluation cannot be performed when training data is unavailable
Solution Approach 1:
The patent creates a pseudo-ground-truth by copying and comparing inference results across multiple noise conditions. Instead of relying on external correct answer data, the system generates reference inference results from noise-free or lightly-noised images and compares them against results from heavily-noised images to determine robustness
Solution Approach 2:
The patent introduces an intermediary reference inference result that mediates between the unknown ground truth and the noisy inference results. This reference result serves as a proxy for correct answers, enabling robustness measurement without direct access to training labels
2Measurement precision
If correct answer data is required for inference, then evaluation is accurate, but the method cannot handle object detection tasks
Solution Approach 1:
The patent designs a universal evaluation framework that handles both classification and object detection tasks through the same core mechanism. The system extracts appropriate metrics (class labels for classification, bounding box coordinates for detection) from inference results and applies consistent robustness measurement logic across different task types
Solution Approach 2:
The patent adapts the evaluation parameters based on the task type. For classification, it measures label consistency; for object detection, it measures bounding box coordinate stability. This parameter adaptation allows accurate robustness measurement tailored to each task's specific output format
Data Source
AI summary
A robustness measurement device of an embodiment includes an acquirer, a noise adder, an inference result group generator, and a robustness measurer. The acquirer acquires an input image. The noise adder adds noise to the input image to generate at least one noise-added image by adding noise to the input image. The inference result group generator generates an inference result group by grouping mutually associated inference results among inference results for at least one of the input image or the noise-added image. The robustness measurer measures robustness of a test target model on the basis of the inference result group.