Frame Selection for Prototype-kNN Patch Attack Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models are vulnerable to patch attacks that degrade accuracy by presenting adversarial patches, and techniques like prototype-kNN struggle to obtain suitable samples due to similar image features or increased processing time.
Innovation Solution
A method involving segmentation of frames to identify regions reflecting objects, calculating overlap and similarity indices, and excluding unsuitable frames to obtain diverse samples for prototype-kNN, improving accuracy and reducing processing time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple samples are prepared using existing techniques, then detection capability against patch attacks is improved, but processing time increases
Solution Approach 1:
The patent extracts only the essential features needed for detection by calculating overlap indices between frames and similarity indices with training data, rather than processing entire frames. This selective extraction of critical information maintains detection reliability while significantly reducing processing time.
Solution Approach 2:
The patent segments the detection process into distinct stages: frame overlap calculation, similarity index calculation against training data, and threshold-based filtering. This segmentation allows each stage to be optimized independently, improving overall processing efficiency while maintaining detection accuracy.
2Measurement precision
If frames with similar image features are used as samples, then detection accuracy is improved, but processing complexity increases
Solution Approach 1:
The patent changes the parameter space by using overlap indices and similarity indices as intermediate representations rather than directly comparing raw image pixels. This parameter transformation simplifies the comparison process while maintaining the ability to detect subtle differences in adversarial patches.
Solution Approach 2:
The patent applies different processing strategies to different types of frames: using overlap indices for chronologically contiguous frames and similarity indices for comparison with training data. This localized approach optimizes processing complexity for each specific comparison scenario while maintaining high detection accuracy.
3Manufacturing precision
If comprehensive frame analysis is performed, then sample suitability is improved, but processing time increases
Solution Approach 1:
The patent performs preliminary filtering by calculating overlap indices between chronologically contiguous frames before conducting more computationally intensive similarity comparisons with training data. This preliminary action eliminates obviously unsuitable frames early, reducing the number of frames requiring full analysis and thus reducing total processing time while maintaining sample suitability.
Solution Approach 2:
The patent applies a two-level analysis: a partial quick check using overlap indices for all frames, and a more comprehensive similarity index calculation only for frames that pass the initial filter. This partial action approach ensures sample suitability for promising candidates while avoiding wasteful processing of clearly unsuitable frames.
Data Source
AI summary
For each pair of two chronologically contiguous frames, an information processing device calculates a first index value representing a degree of overlap between regions each identified from a corresponding one of the paired two frames. For each pair of two sequentially contiguous frames, the information processing device calculates a second index value representing a degree of similarity between the paired two frames with respect to a predetermined type of feature. To identify remaining frames, the information processing device, excludes from a series of frames, a certain number of frames anterior and posterior to a pair at least whose first index value is less than a first threshold value or whose second index value is less than a second threshold value. The information processing device extracts a frame whose degree of similarity in the predetermined type of feature to another frame among the remaining frames satisfies a predetermined condition.