在脑瘤分类中,使用集体对抗训练和特征挤压来进行对抗性攻击的多层防御
Ahmeed Yinusa1, Misa Faezipour2
1Computational and Data Science Program, Middle Tennessee State University, 1301 East Main Street, Murfreesboro, TN, 37132, USA.
用于脑瘤分类的深度学习模型容易受到对抗性攻击. 结合对抗训练和功能挤压的防御策略,改善了对常见攻击的模型弹性,提高了医学成像中的AI可靠性.
科学领域:
- 人工智能的人工智能
- 医疗成像医学成像
- 计算神经科学是一种神经科学.
背景情况:
- 深度学习,特别是卷积神经网络 (CNN),在医学图像的脑瘤分类方面显示出前景.
- 然而,这些人工智能模型容易受到敌对攻击,这可能会破坏它们的临床可靠性.
- 对此类攻击的脆弱性对AI在医疗保健中的安全部署构成了重大挑战.
研究的目的:
- 评估基于VGG16的CNN模型对抗对抗攻击的脑瘤分类的稳定性.
- 开发和评估一个多层次的防御战略,以提高模型的弹性.
- 调查防御机制对在对抗条件下的AI模型性能的影响.
主要方法:
- 一个VGG16CNN模型被训练在干净的磁共振成像 (MRI) 数据上进行脑瘤分类.
- 模型的性能在暴露于快速梯度标志方法 (FGSM) 和预测梯度下降 (PGD) 敌对攻击后进行了评估.
- 实施了一种涉及对抗训练 (使用FGSM/PGD示例) 和特征挤压 (比特深度减少,高斯模糊) 的防御策略.
主要成果:
- 在清洁的MRI数据上,VGG16模型实现了96%的准确性.
- 敌对攻击显著降低了准确度,达到32% (FGSM) 和13% (PGD).
- 实施的防御策略提高了准确性,达到54% (FGSM) 和47% (PGD) 对抗对手的例子.
结论:
- 医疗图像分析的深度学习模型虽然准确,但需要强大的防御机制来抵御敌对攻击.
- 积极的防御策略对于确保AI在医学诊断中的可靠性和临床适用性至关重要.
- 这项研究表明,结合对抗训练和特征挤压在提高医疗成像中的AI模型弹性方面的有效性.
更多相关视频
12:50Lesion Explorer: A Video-guided, Standardized Protocol for Accurate and Reliable MRI-derived Volumetrics in Alzheimer's Disease and Normal Elderly
Published on: April 14, 2014
09:53Quantifying the Brain Metastatic Tumor Micro-Environment using an Organ-On-A Chip 3D Model, Machine Learning, and Confocal Tomography
Published on: August 16, 2020
相关概念视频
Cancer Therapies
However, cancer treatments can pose several challenges, as therapies used to kill cancer cells are generally also toxic to normal cells. Moreover, cancer cells mutate rapidly and can develop resistance to chemical agents or radiation therapy. Besides, all types of cancer cells may not respond to the same therapy. Some cancer cells respond to one...
Targeted Cancer Therapies
There are several types of targeted therapies against specific...
Treatment Resistant Cancers
Combination Therapies and Personalized Medicine
The combination of the drug acetazolamide and sulforaphane is a good example of combination therapy to treat cancer. The cells in the interior of a large tumor often die due to the hypoxic and...
Treatment Resistent Cancers
Tumor Immunotherapy
