Automatic Kidney and Tumor Segmentation with Attention-based V-Net

Hu, Yucheng; Deng, Han; Zhou, Yang; Chen, Yimin; Zhou, Hao; Yang, Wanqi

doi:10.24926/548719.087

Cited by 1 publication

(2 citation statements)

References 3 publications

(3 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The visualization of synthesized OOD images. We mark the OOD signals with red boxes, which are "lesions" generated by random shapes and textures [59,60]. "Tiny, Small, Medium, and Large" are the sizes of lesions, while the numbers are the quantities of the lesions.…”

Section: G Comparison 3d and 2d Vulnerabilitymentioning

confidence: 99%

“…In this section, we try to evaluate the possibility of using HFC term to detect the out-of-distribution (OOD) outlier data. Because obtaining the real-world OOD signals (e.g., artifacts, anatomical variations, or unseen pathologies) is difficult, we modify lesion synthesis methods proposed by Hu et al [59,60] to simulate OOD images by synthesizing different sizes and quantities of "lesions" and insert into a random position in clean images, which are served as anomaly signals and shown in Fig. 8.…”

Section: H Detecting Ood Signals Using Hfc Termmentioning

confidence: 99%

See 1 more Smart Citation

A Hierarchical Feature Constraint to Camouflage Medical Adversarial Attacks

Yao

Lin

et al. 2021

Medical Image Computing and Computer Assisted Intervention – MICCAI 2021

View full text Add to dashboard Cite

Deep learning based methods for medical images can be easily compromised by adversarial examples (AEs), posing a great security flaw in clinical decisionmaking. It has been discovered that conventional adversarial attacks like PGD which optimize the classification logits, are easy to distinguish in the feature space, resulting in accurate reactive defenses. To better understand this phenomenon and reassess the reliability of the reactive defenses for medical AEs, we thoroughly investigate the characteristic of conventional medical AEs. Specifically, we first theoretically prove that conventional adversarial attacks change the outputs by continuously optimizing vulnerable features in a fixed direction, thereby leading to outlier representations in the feature space. Then, a stress test is conducted to reveal the vulnerability of medical images, by comparing with natural images. Interestingly, this vulnerability is a double-edged sword, which can be exploited to hide AEs. We then propose a simple-yet-effective hierarchical feature constraint (HFC), a novel add-on to conventional white-box attacks, which assists to hide the adversarial feature in the target feature distribution. The proposed method is evaluated on three medical datasets, both 2D and 3D, with different modalities. The experimental results demonstrate the superiority of HFC, i.e., it bypasses an array of state-of-the-art adversarial medical AE detectors more efficiently than competing adaptive attacks 1 , which reveals the deficiencies of medical reactive defense and allows to develop more robust defenses in future. Index Terms-Security in Machine Learning, adversarial attacks and defense I. INTRODUCTION D EEP neural networks (DNNs) are vulnerable to adversarial examples (AEs) [1], which are maliciously computed by adding human-imperceptible perturbations to clean images.

show abstract

Section: G Comparison 3d and 2d Vulnerabilitymentioning

confidence: 99%

Section: H Detecting Ood Signals Using Hfc Termmentioning

confidence: 99%