Abstract: In recent years,brain functional networks analysis using the resting state functional magnetic resonance imaging data has been widely used in computer-aided diagnosis tasks of various brain diseases.The graph convolutional network framework integrating clinical phenotypic measurements and brain functional networks improves the applicability of intelligent medical disease diagnosis models to the real world.However,the trustworthiness study is an important but still widely neglected component of disease diagnosis models based on brain functional networks.Adversarial attack techniques in medical machine learning can deceive models,which further leads to the security and trust issues of the model applied in clinical practice.Based on this,this paper proposes an adversarial attack method BFGCNattack on graph convolutional network for brain disease diagnosis,constructs a disease diagnosis model integrating clinical phenotypic measurements,and evaluates the robustness of brain functional networks-based disease diagnosis model in the face of adversarial attacks.Experimental results on the autism brain imaging data exchange dataset suggest that the models constructed using graph convolutional networks are vulnerable to the proposed adversarial attack.Even if only a small number(10%) of perturbations are performed,the model’s accuracy and classification margin significantly decrease,while the fooling rate significantly increases.

Key words: Adversarial attack method, Brain disease diagnosis, Graph convolutional network, Brain functional networks analysis, Model robustness