Digital Watermarking as an Adversarial Attack on Medical Image Analysis with Deep Learning

Area of Science:

• Computer Vision
• Medical Image Analysis
• Deep Learning

Background:

• Deep Neural Networks (DNNs) are widely used in Computer Vision (CV).
• Adversarial attacks pose a significant threat to the performance of CV models, especially in Medical Image Analysis.
• Digital watermarking is typically used for security but can be repurposed.

Purpose of the Study:

• To investigate digital watermarking as a novel black-box adversarial attack, termed 'watermarking attacks'.
• To assess the impact of these watermarking attacks on the performance of state-of-the-art CV models used for medical image analysis.
• To identify the most vulnerable models and medical imaging modalities.

Main Methods:

• A moment-based local image watermarking method was implemented.
• The method was applied to three medical imaging modalities: Magnetic Resonance Images (MRI), Computed Tomography (CT-scans), and X-ray images.
• The attacks were tested against three CV models: DenseNet 201, DenseNet 169, and MobileNetV2.

Main Results:

• The watermarking attacks caused over 50% degradation in model accuracy.
• MobileNetV2 demonstrated the highest vulnerability to the proposed attacks.
• CT-scans experienced the most significant performance reduction among the tested modalities.

Conclusions:

• Digital watermarking presents a viable and potent adversarial attack vector against CV models in medical imaging.
• The widespread use of watermarks for security may inadvertently introduce substantial risks to AI-driven vision systems.
• Further research is needed to develop robust defenses against such watermarking attacks.