Adversarial Noise Removal in Image Recognition via Resolution Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Adversarial sample attacks, which introduce imperceptible noise, significantly affect the accuracy of neural-network-based image recognition models, leading to potential misjudgments in critical applications like autonomous driving, where incorrect decisions can result from undetected disturbances in image data.
Innovation Solution
An image processing method that reduces noise by performing operations such as reducing resolution or smoothing on initial images with adversarial noise, followed by training an image enhancement model using sample and reference images to generate a target image with higher quality, thereby mitigating the impact of adversarial attacks while maintaining image recognition capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If image processing operations (resolution reduction or smoothing) are performed on initial images with adversarial noise, then the noise associated with adversarial sample attacks is reduced, but the image quality deteriorates
Solution Approach 1:
The patent applies image processing operations (resolution reduction or smoothing) as a preliminary step before recognition to remove adversarial noise. This preliminary action modifies the input image to eliminate harmful factors before the recognition model processes it, thereby protecting against adversarial attacks while preserving essential image information for accurate recognition.
2Reliability
If image processing operations are performed to reduce adversarial noise, then the reliability of image recognition improves, but the detail information of the image is lost
Solution Approach 1:
The patent applies different image processing operations to different regions or aspects of the image based on local needs. Smoothing operations are applied to regions with adversarial noise while preserving important detail regions, and resolution reduction is selectively applied. This local differentiation allows the system to improve reliability by removing noise while minimizing information loss in critical areas.
Data Source
AI summary
According to embodiments of the present disclosure, a method and an apparatus for processing an image, a device, and a storage medium are provided. The method includes: performing an image processing operation on an initial image having a noise associated with an adversarial sample attack, to obtain an intermediate image, the image processing operation including at least one of: reducing resolution of the initial image, or smoothing at least a part of the initial image; determining an image enhancement model matching the image processing operation, the image enhancement model being trained based on a sample image and a reference image, and the reference image being obtained by performing at least the image processing operation on the sample image; and generating a target image by processing the intermediate image using the image enhancement model, the target image having an image quality higher than the intermediate image.