Object Detection Adversarial Patch Screening via Object-Size Missingness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting adversarial patch attacks in image AI for object detection, such as at self-checkouts and license plate recognition, are ineffective when missingness with random size and position fails to cover the patch, leading to overlooked fraud and misrecognition.
Innovation Solution
A computer-readable recording medium that implements an information processing program to determine a missing rate based on the minimum size of an adversarial patch, adding missingness exceeding this rate to an image, and comparing detection results to identify and alert on adversarial patch usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If missingness with random size and position is added to detect adversarial patches, then the detection method can identify some patches, but it fails when the random missingness does not cover the patch, leading to missed detection
Solution Approach 1:
The patent changes the parameters of missingness from random size and position to a systematic approach where missingness size is dynamically adjusted based on the detected object's size. The missingness is added at specific locations (center, corners, edges) rather than randomly, ensuring coverage of potential adversarial patches while maintaining computational efficiency and detection reliability.
2Productivity
If the missingness addition method is simple and fast, then processing time is reduced, but the detection accuracy against adversarial patches deteriorates
Solution Approach 1:
The patent segments the image processing into distinct stages: object detection, object size determination, missingness rate calculation, and selective missingness addition. By segmenting the process and only applying missingness to regions where objects are detected, the system maintains high processing speed while improving detection accuracy against adversarial patches.
Solution Approach 2:
The patent applies local quality by adding missingness at specific locations (center, corners, edges) of detected objects rather than uniformly across the entire image. This localized approach ensures that missingness is added where adversarial patches are most likely to appear, improving detection accuracy without the computational cost of global processing.
Data Source
AI summary
A non-transitory computer-readable recording medium stores therein an information processing program that causes a computer to execute a process including acquiring a first value representing a size of a region of an object included in a first image, determining a missing rate for an adversarial patch, based on a second value representing a minimum size of the adversarial patch acquired according to the first value, generating a second image in which missingness exceeding the missing rate is added to the first image, and comparing a first detection result obtained by inputting the first image into an object detection model with a second detection result obtained by inputting the second image into the object detection model.