Projection Area Analysis for Adversarial Patch Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection models are vulnerable to adversarial patches, leading to inaccurate detections, especially in environments where training data with adversarial patches is not readily available.
Innovation Solution
An information processing apparatus that combines an object detection model, an object extraction model, and a projection area extraction model to detect adversarial patches without prior training, using a combination of classification results from these models and image processing techniques to enhance detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If training data including adversarial patches is prepared and training is performed to classify scenes with adversarial patches into a new class, then detection of adversarial patches is enabled, but it becomes difficult to detect adversarial patches when new items are added to inventory daily due to inability to prepare comprehensive training data
Solution Approach 1:
The detection process is segmented into two independent components: (1) an object detection model that identifies objects in scenes, and (2) a projection area extraction model that extracts geometric projection areas of detected objects. This segmentation allows the system to detect adversarial patches without requiring training data for each new item, as the projection area extraction is based on geometric properties rather than learned features specific to particular items.
Solution Approach 2:
The projection area extraction model serves multiple functions: it extracts geometric projection areas for any detected object regardless of item type, enables adversarial patch detection through threshold comparison, and adapts to new items immediately without retraining. This universal approach allows the same model to handle both traditional objects and newly added items equally effectively.
2Productivity
If traditional object detection models are used, then object detection is performed, but the models are vulnerable to adversarial patches leading to inaccurate detection
Solution Approach 1:
A projection area extraction model is introduced as an intermediary component between the object detection model and the final detection result. This intermediary extracts geometric projection areas of detected objects and compares them against threshold values to determine whether adversarial patches are present. The intermediary adds a verification layer that protects against adversarial attacks while maintaining the speed of the original object detection model.
Solution Approach 2:
The system changes the parameter being analyzed from learned object features to geometric projection area parameters. By extracting and analyzing the projection area (a geometric parameter) rather than relying solely on learned features that can be manipulated by adversarial patches, the system achieves more reliable detection. The projection area parameters are compared against threshold values to make detection decisions.
Data Source
AI summary
An information processing apparatus extracts a projection area of an object from an input scene containing a projection of the object, specifies each of second areas that are included in the extracted projection area of the object and that are similar to a plurality of first areas representing a feature of each of labels in a feature space, outputs a first classification result based on each of the plurality of labels indicating a state in which a distribution of combinations of the specified second areas and the first areas that are associated with the second areas is closer than a predetermined threshold; and determines whether or not an adversarial patch is included in the object by comparing the first classification result with a second classification result that is a result of classification of the objects obtained by inputting the input scene to a predetermined object detection model.