.
With the increasing use of artificial intelligence (AI) in medicine, medical image analysis systems have become targets for cyber attacks, compromising diagnoses and putting patients’ health at risk. To face this challenge, PhD student Erikson Júlio de Aguiar, from the Institute of Mathematics and Computational Sciences (ICMC), developed Radar-Mix, a tool that detects and analyses external attacks on AI systems. The article that describes how the solution was created and the results achieved entitled Radar-Mix: How to Uncover Adversarial Attacks in Medical Image Analysis through Explainability, was awarded Best Student Paper at the 37th International Symposium on Computer-Based Medical Systems(CBMS), held from June 26 to June 28 in Guadalajara, Mexico.
According to Erikson, hacker attacks insert small alterations or noise into medical images, such as those obtained by X-ray and MRI, to confuse systems and lead to misdiagnoses. Radar-Mix was created precisely to mitigate this type of threat, providing a detailed analysis of attacks.“Artificial intelligence has been a great ally in diagnosing diseases, but at the same time, hackers have found ways to compromise these systems. Radar-Mix is a tool that makes it possible not only to detect these attacks but also to understand how and where they happen, which is essential for ensuring the security of diagnoses,” explains Erikson.
The model uses advanced tools such as Grad-CAM and Shap, which help to visually show which parts of an image have been altered by an attack. This detail makes protection more efficient, as Radar-Mix not only detects whether the system has been attacked but also explains where and how the attack happened, offering a more complete analysis compared to other existing solutions.
International Recognition
The Best Student Paper award at CBMS 2024 is a recognition of the work carried out by Erikson under the guidance of Professors Agma Traina, his doctoral advisor, and Caetano Traina, both from ICMC. “Seeing Erikson’s work be awarded at one of the world’s leading computer and medical forums shows the quality of what we are developing here at ICMC,” says Professor Agma.
According to the professor, Radar-Mix has the potential to be implemented directly in medical systems, protecting hospitals and diagnostic centers against cyber attacks that could compromise patient treatment. “Radar-Mix was developed in a modular way, facilitating its implementation and adaptation to different contexts. The interposition has been structured as a filter between access to services and servers, offering an additional layer of security and control”, explains the Professor.
This protection is essential, especially in a context where the number of cyber attacks on healthcare institutions has been growing: “In the United States, for example, an attack can affect a hospital’s reputation and lead to major financial and operational losses. Radar-Mix helps to maintain trust in these AI systems, which is crucial for both doctors and patients“, Erikson added.
Radar-Mix is available for use, with its open-source code available for reproduction in other research or for improvements. Those interested may access and contribute via this link.
Next steps
Being recognized at CBMS 2024, Erikson continues to expand his research and has developed a new paper, showing a new perspective on Radar-Mix, which will be presented at Spie Medical Imaging(International Society for Optics and Photonics) in February 2025 in California, United States. In addition, the PhD student is working with specialists from the University of Florida, where he is doing a one-year sandwich internship, funded by the São Paulo Research Foundation (Fapesp).
Professor Agma Traina also emphasizes the importance of this international experience: “Erikson is working with big names in the field of AI safety and this experience abroad will contribute a lot to his career, as he is an extremely dedicated student. I’m sure that this award is just the start of a brilliant academic career”.
Access the winning paper by clicking here. Access the project’s GitHub..
*Text from the ICMC Communications Office
English version: Nexus Traduções
