ROSA: Robust Salient Object Detection Against Adversarial Attacks.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Deep convolutional neural networks (CNNs) and fully convolutional networks (FCNs) have significantly advanced salient object detection.
• Current FCN-based methods achieve high accuracy but are susceptible to adversarial attacks, where subtle noise can mislead models without altering ground truth.

Purpose of the Study:

• To investigate the effectiveness of adversarial attacks on state-of-the-art salient object detection models.
• To propose and validate a novel, end-to-end trainable framework for enhancing the robustness of FCN-based salient object detection against adversarial perturbations.

Main Methods:

• The study first demonstrates successful adversarial attacks on existing salient object detection models.
• A new framework is introduced, incorporating a segment-wise shielding component to destroy adversarial noise and a context-aware restoration component for saliency map refinement.
• The proposed method is end-to-end trainable and designed to work with arbitrary FCN-based models.

Main Results:

• Adversarial samples were confirmed to be effective against a wide range of existing salient object detection methods.
• The proposed robustness framework significantly improved the performance of state-of-the-art models when tested on multiple datasets.
• The segment-wise shielding and context-aware restoration components effectively preserved image boundaries and refined saliency maps.

Conclusions:

• Salient object detection models are vulnerable to adversarial attacks, necessitating robust defense mechanisms.
• The proposed novel framework offers an effective solution for enhancing the adversarial robustness of FCN-based salient object detection models.
• This research contributes a significant advancement in securing deep learning models for computer vision tasks against malicious perturbations.