Identify Vulnerability of Adversarial Attack on Chest X-Ray Image Using Hybrid Refinement - Generative Adversarial Network with Convolutional Block Attention Module

Chest X-ray (CXR) imaging is the most widely applied diagnostic tool in healthcare, and plays a vital role in radiological evaluations.However, machine learning techniques such as CXR analysis applied to medical imaging face significant security threats from adversarial attacks.These attacks exploit system vulnerabilities, compromising diagnostic accuracy and reliability by altering or misleading diagnostic images.The challenges in analyzing adversarial attacks effects on CXR images include struggles in distinguishing diagnostic features from unwanted noises.The proposed CXR specific Hybrid Refinement -Generative Adversarial Network with Convolutional Block Attention Module (CXR HRGAN-CBAM) efficiently detects adversarial attacks, refines affected image areas, and optimally enhances both channel and spatial features.Initially, raw data obtained from the NIH CXR dataset, after which preprocessing is carried out with Contrast Limited Adaptive Histogram Equalization (CLAHE) to improve image quality and identify the visibility of small regions in CXR images.The proposed method achieves an accuracy of 99.12% on the NIH CXR dataset, outperforming existing methods such as Self-Attention Generative Adversarial Capsule Network optimized with Sun Flower Optimization algorithm (GACaps-SFO).