Improving Robustness of Medical Image Diagnosis System by Using Multi-loss Hybrid Adversarial Function with Heuristic Projection

Abstract

AbstractMedical image diagnosis system by using deep neural networks (DNN) can improve the sensitivity and speed of interpretation of chest CT for COVID-19 screening. However, DNN based medical image diagnosis is known to be influenced by the adversarial perturbations. In order to improve the robustness of medical image diagnosis system, this paper proposes an adversarial attack training method by using multi-loss hybrid adversarial function with heuristic projection. Firstly, the effective adversarial attacks which contain the noise style that can puzzle the network are created with a multi-loss hybrid adversarial function (MLAdv). Then, instead of adding these adversarial attacks to the training data directly, we consider the similarity between the original samples and adversarial attacks by using an adjacent loss during the training process, which can improve the robustness and the generalization of the network for unanticipated noise perturbations. Experiments are finished on COVID-19 dataset. The average attack success rate of this method for three DNN based medical image diagnosis systems is 63.9%, indicating that the created adversarial attack has strong attack transferability and can puzzle the network effectively. In addition, with the adversarial attack training, the augmented networks by using adversarial attacks can improve the diagnosis accuracy by 4.75%. Therefore, the augmented network based on MLAdv adversarial attacks can improve the robustness of medical image diagnosis system.KeywordsMedical image diagnosis systemAdversarial attackMulti-loss hybrid functionHeuristic projectionAttack transferability