Authors: Pranav Kulkarni, Andrew Chan, Nithya Navarathna, Skylar Chan, Paul H. Yi, Vishwa S. Parekh

Published on: February 08, 2024

Impact Score: 8.6

Arxiv code: Arxiv:2402.05713

Summary

• What is new: Demonstration of demographically targeted label poisoning attacks introducing adversarial underdiagnosis bias in DL models in radiology without impacting overall performance.
• Why this is important: Deep learning models in radiology might exacerbate clinical biases against vulnerable patient populations.
• What the research proposes: Investigating and highlighting the risks of demographically targeted adversarial attacks on DL models in a clinical environment.
• Results: Adversarial underdiagnosis bias can be introduced in models, affecting performance on underrepresented groups without affecting overall model metrics.

Technical Details

Technological frameworks used: nan

Models used: Deep learning models for radiology.

Data used: Demographically diverse datasets, focused on sex, age, and intersectional subgroups.

Potential Impact

Healthcare, particularly radiology departments and AI software providers in medical imaging.

Want to implement this idea in a business?

We have generated a startup concept here: EquiScan.