Universal adversarial attacks on deep neural networks for medical image classification

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Cybersecurity

Background:

• Deep neural networks (DNNs) are crucial for automated clinical diagnosis via medical image classification.
• Evaluating DNN robustness against adversarial attacks is vital due to high-stakes medical decisions.
• Previous research on simple adversarial attacks is insufficient; vulnerability to Universal Adversarial Perturbations (UAPs) remains underexplored.

Purpose of the Study:

• To investigate the vulnerability of DNNs in medical image classification to Universal Adversarial Perturbations (UAPs).
• To assess DNN performance across three distinct medical classification tasks: skin cancer, diabetic retinopathy, and pneumonia.
• To evaluate the impact of UAPs on seven different DNN model architectures.

Main Methods:

• Focused on three representative DNN-based medical image classification tasks.
• Investigated vulnerability against seven model architectures of Universal Adversarial Perturbations (UAPs).

Main Results:

• DNNs demonstrated significant vulnerability to both non-targeted and targeted UAPs, achieving over 80% success rates.
• Model architecture had minimal impact on DNN vulnerability to UAPs.
• Adversarial retraining proved effective in only a few cases for enhancing robustness against UAPs.

Conclusions:

• DNN-based clinical diagnosis systems are more susceptible to adversarial attacks than previously assumed.
• Adversaries can compromise diagnostic accuracy with lower costs and without data distribution considerations.
• Current adversarial defense strategies may have limitations, necessitating careful development and application of DNNs in medical imaging.