Abstract
The classification of medical images is an essential component in healthcare, serving as a cornerstone for accurate diagnosis and effective treatment planning. Deep neural network (DNN)-based methods have demonstrated significant potential in this domain, delivering exceptional outcomes by leveraging advanced computational power to recognize complex patterns in medical images. However, despite these advancements, DNNs are susceptible to adversarial attacks—deliberate perturbations in input data designed to deceive the model, leading to misclassifications. These attacks can significantly impair the performance of DNNs and undermine their reliability, posing a substantial risk in critical healthcare applications where accuracy is paramount. Current defense strategies against adversarial attacks, while somewhat effective in general machine learning contexts, often fall short when applied to medical image analysis. This shortfall is largely due to the inherent characteristics of medical images, such as high variability, the presence of noise, and the necessity for high precision. Medical images are typically more complex and less abundant than natural images, making it challenging to apply standard adversarial defense mechanisms effectively. These unique challenges necessitate the development of specialized defense strategies tailored to the nuances of medical imaging. This research focuses on designing new approaches and addressing the above challenges by •improving the robustness of deep learning models, especially in the classification of medical ultrasound imaging for breast cancer especially by using methodology based on Robust Self Training,
•improve models’ generalization ability and solve overfitting problem for model by implementing pretrained weights and testing the impact of it on generalization,
•build new pooling layers to improve model robustness, and
•improve robust self-training technique to make more accurate robust model.
In summary, our research contributes to the ongoing efforts to secure deep learning models in healthcare by developing a defense framework that addresses the specific challenges posed by medical image analysis. This work not only enhances the robustness of DNNs but also paves the way for more trustworthy artificial intelligence (AI) applications in critical healthcare environments, ultimately aiming to improve patient outcomes and bolster confidence in AI-assisted diagnostic tools.