Blessing or a Curse. Discussing Security Concerns of Diagnostic Models in Radiological Assessment

NLDL 2025 Conference Submission10 Authors

Keywords: Deep Learning, computer vision, radiology

TL;DR: Data poisoning remains a threat to medical AI with possible solutions found in fraud detection.

Abstract:

Radiology is increasingly adopting AI-based workflows, which provide promise but also introduce new security concerns. The goal of this research is to enhance the security of these workflows by evaluating the risks of data poisoning attacks using the Fast Gradient Sign Method (FGSM) and Carlini-Wagner (C&W) techniques. The dataset utilized is from the 2017 RSNA Pediatric Bone Age Challenge. Detection methods commonly employed in financial fraud are evaluated to assess their effectiveness in this context. Knowledge distillation will also be explored as a defense mechanism against data poisoning, offering a potential mitigation strategy. By conducting these evaluations and proposing defenses, this research aims to contribute to the more robust deployment of AI systems in real-world radiology applications.

Submission Number: 10