Medical digital twins—virtual models replicating human biological systems—have shown promise in predicting diseases with high precision. However, they are susceptible to cyberattacks that can manipulate data, leading to incorrect diagnoses and jeopardizing patient safety.

A team of researchers from Dongguk University, South Korea, in collaboration with Oregon State University, USA, says it has developed a breakthrough solution: Wavelet-Based Adversarial Training (WBAD). This new defense system, the researchers say, restores the accuracy of medical digital twins to 98%, even when subjected to sophisticated adversarial attacks.

Addressing vulnerabilities in medical digital twins

A digital twin is a real-time virtual representation of a physical system, used to simulate, test, and optimize performance. In healthcare, medical digital twins replicate biological systems, enabling the prediction of diseases and the testing of medical treatments. However, as mentioned, these models are vulnerable to adversarial attacks, where slight, deliberate alterations to input data can mislead the system into providing false predictions—such as incorrect cancer diagnoses—putting patients at risk.

To address this issue, the research team, led by Professor Insoo Sohn of Dongguk University, proposed the innovative WBAD defense system, a dual-layer approach combining wavelet denoising and adversarial training. Their findings were first published in Information Fusion in March 2025, following an initial online release in October 2024.

A dual-layer approach to security

Professor Sohn explains, “This is the first study to propose a secure medical digital twin system using a two-stage defense mechanism that integrates wavelet denoising and adversarial training to safeguard against cyberattacks.”

The team applied WBAD to a breast cancer diagnostic model utilizing thermography images, which detect temperature variations in the body. These images are particularly useful for identifying tumors, which appear as hotter areas due to increased blood flow and metabolic activity. The model processes these thermographic images using Discrete Wavelet Transform (DWT) to extract key features, which are then used in a machine learning classifier to differentiate between healthy and cancerous tissue.

From 92% to 5% accuracy: The impact of adversarial attacks

Initially, the breast cancer diagnostic model achieved an accuracy rate of 92%. However, when subjected to three common types of adversarial attacks—Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Carlini & Wagner (C&W) attacks—accuracy plummeted to just 5%, highlighting the vulnerability of digital twins to manipulation.

To address this, the researchers introduced their two-tier defense. The first stage, wavelet denoising, is applied during image preprocessing. Adversarial attacks typically inject high-frequency noise into input data to disrupt the model’s performance. The wavelet denoising process removes this noise while preserving the image’s essential features.

The second layer involves adversarial training, where the model learns to recognize and counter adversarial inputs. This combined strategy proved highly effective, restoring accuracy to 98% for FGSM attacks, 93% for PGD attacks, and 90% for C&W attacks.

A new standard in medical digital twin security

“Our findings mark a significant step forward in the security of medical digital twins,” said Professor Sohn.

“This dual defense mechanism not only protects against cyberattacks but also enhances the functionality and reliability of medical diagnostic systems.”

This approach offers a promising solution for securing medical digital twins, ensuring they can be relied upon to deliver accurate diagnoses and improve patient care, even in the face of cyber threats. The research represents a moment of change in the intersection of cybersecurity and healthcare innovation, setting a new standard for the protection of these powerful tools.