Salient object detection dataset with adversarial attacks for genetic programming and neural networks.

Area of Science:

• Computer Vision
• Machine Learning Security
• Deep Learning

Background:

• Machine learning (ML) excels in feature extraction but faces security challenges due to adversarial perturbations.
• Salient object detection (SOD) using deep convolutional neural networks (CNNs) is effective but vulnerable to attacks, impacting trustworthiness.

Purpose of the Study:

• To create a comprehensive dataset for evaluating the adversarial robustness of salient object detection models.
• To provide a resource for analyzing and developing defenses against adversarial attacks in computer vision.

Main Methods:

• Compiled a dataset of 56,387 digital images by applying 12 types of adversarial examples to five distinct image databases.
• Included standard SOD databases (FT, PASCAL-S, ImgSal, DUTS) and a real-world visual attention dataset (SNPL).
• Provided original and rescaled images for accessibility and distribution.

Main Results:

• The dataset facilitates the assessment of model performance under various adversarial attack scenarios.
• It enables quantitative analysis of the impact of different adversarial perturbations on SOD models.

Conclusions:

• The developed dataset is crucial for advancing research in robust salient object detection.
• It serves as a valuable resource for the machine learning community to address security vulnerabilities in computer vision systems.