Transformative Defense System for Medical Digital Twins Developed by Researchers

Enhancing Safety in Medical Digital Twins with WBAD

In a groundbreaking development, researchers at Dongguk University have created a state-of-the-art defense system, known as Wavelet-Based Adversarial Training (WBAD), designed specifically for medical digital twins. Digital twins, which are virtual replicas of biological systems, hold immense potential in predicting diseases and enhancing healthcare outcomes. However, their susceptibility to cyberattacks poses significant risks, particularly concerning patient safety and diagnostic accuracy.

Medical digital twins use real-time data to create accurate simulations of a person's biological processes. They enable healthcare professionals to rigorously test treatments, predict diseases, and study various treatment outcomes without the need for invasive procedures. Unfortunately, these powerful models can be vulnerable to adversarial attacks—where tiny alterations to input data lead to erroneous diagnoses, such as incorrect cancer predictions.

To counteract these vulnerabilities, the research team, led by Professor Insoo Sohn, developed WBAD, which employs a pioneering two-stage defense mechanism. The first component of this mechanism integrates wavelet denoising methods to eliminate high-frequency noise from input images—noise that is often introduced by adversarial attacks. The second component incorporates adversarial training, ensuring that the machine learning model is attuned to recognize and resist these malicious inputs.

Initially, the model used for breast cancer detection achieved a promising accuracy of 92% when making predictions based on thermography images. This technique analyzes body temperature variations, as tumors commonly appear as warmer areas due to increased metabolic activity. However, when exposed to various types of adversarial attacks, the model's accuracy plummeted alarmingly to a mere 5%, highlighting serious safety concerns.

To address this, Professor Sohn and his team implemented wavelet denoising during the preprocessing stage of image input. Soft thresholding was applied, effectively removing unwanted noise while retaining the vital low-frequency features necessary for accurate diagnosis. This crucial first step ensures that the model is not misled by adversarial inputs.

Moreover, the incorporation of adversarial training allows the model to identify and resist attempts at manipulation. By continually retraining the model, it acclimatizes to adversarial tactics, improving its robustness against these threats. After implementing the two-layer defense strategy of WBAD, the model's accuracy soared back to 98% against Fast Gradient Sign Method (FGSM) attacks, while also maintaining an impressive 93% and 90% accuracy against Projected Gradient Descent (PGD) and Carlini-Wagner (CW) attacks, respectively.

Professor Sohn states, “Our findings represent a major leap in ensuring the security of medical digital twins. By providing a comprehensive defense framework against cyber threats, we can inherently enhance the functionality and reliability of these systems.” This advancement promises to bolster patient safety in healthcare technology, setting a new standard for digital frameworks used in medical diagnostics.

In conclusion, the Wavelet-Based Adversarial Training established by Dongguk University represents not only a significant contribution to the field of Digital Twin Security but also a vital step forward in the evolution of healthcare technology. It ensures that these advanced systems can deliver reliable results, ultimately improving the health outcomes of patients by harnessing the full potential of modern medical technology. The research findings are published in the March 2025 issue of the journal Information Fusion, further highlighting the academic implications of this important discovery.

For more information, please visit the Dongguk University website.