Master's Thesis · 2026

MedDef: An Attention Based Model For Adversarial Resilience in Medical Imaging

A dual-frequency attention mechanism (DAAM) for defending medical image classifiers against adversarial attacks across 3 datasets and 8 attack types.

Enoch Kwateh Dongbo

东博

202324100003
enoch.dongbo@stu.ujn.edu.cnekd@ekddigital.com
+86 185 0683 2159

Supervisor

Niu Sijie

牛四杰

3

Datasets

6

Model Variants

8

Attack Types

DAAM

Defense

Supported Datasets

Tuberculosis Chest X-Ray

Binary classification: Normal vs Tuberculosis chest X-rays

Normal
Tuberculosis

Chest X-Ray (Pneumonia)

Binary classification: Normal vs Pneumonia chest X-rays

NORMAL
PNEUMONIA

Retinal OCT

4-class retinal OCT classification: CNV, DME, Drusen, Normal

CNV
DME
DRUSEN
NORMAL