Universal adversarial attacks on deep neural networks for medical image classification

摘要

Deep neural networks (DNNs) are widely investigated in medical image classification to achieve automated support for clinical diagnosis. It is necessary to evaluate the robustness of medical DNN tasks against adversarial attacks, as high-stake decision-making will be made based on the diagnosis. Several previous studies have considered simple adversarial attacks. However, the vulnerability of DNNs to more realistic and higher risk attacks, such as universal adversarial perturbation (UAP), which is a single perturbation that can induce DNN failure in most classification tasks has not been evaluated yet. We focus on three representative DNN-based medical image classification tasks (i.e., skin cancer, referable diabetic retinopathy, and pneumonia classifications) and investigate their vulnerability to the seven model architectures of UAPs. We demonstrate that DNNs are vulnerable to both nontargeted UAPs, which cause a task failure resulting in an input being assigned an incorrect class, and to targeted UAPs, which cause the DNN to classify an input into a specific class. The almost imperceptible UAPs achieved??80% success rates for nontargeted and targeted attacks. The vulnerability to UAPs depended very little on the model architecture. Moreover, we discovered that adversarial retraining, which is known to be an effective method for adversarial defenses, increased DNNs’ robustness against UAPs in only very few cases. Unlike previous assumptions, the results indicate that DNN-based clinical diagnosis is easier to deceive because of adversarial attacks. Adversaries can cause failed diagnoses at lower costs (e.g., without consideration of data distribution); moreover, they can affect the diagnosis. The effects of adversarial defenses may not be limited. Our findings emphasize that more careful consideration is required in developing DNNs for medical imaging and their practical applications.