Table 3 IoU, overlap difference (OD), trigger detection rate (TDR) results of generated using seven XAI models during a poisoning attack with dynamic triggers for detection effectiveness evaluation.

From: Improving explainable AI with patch perturbation-based evaluation pipeline: a COVID-19 X-ray image analysis case study

Trigger size

Evaluation

BP

Guided BP

GradCAM

Guided GradCAM

OS

Ablation

LIME

20 * 20

IoU

0.4810

0.6994

0.0280

0.2662

0.4261

0.4653

0.1527

OD

0.0018

0.0013

0.0056

0.0036

0.0023

0.0027

0.0004

TDR

0.9205

0.9323

0.2739

0.5405

0.8233

0.8115

0.9764

40 * 40

IoU

0.4481

0.6590

0.1597

0.4676

0.3421

0.3379

0.4778

OD

0.0084

0.0074

0.0226

0.0121

0.0112

0.0195

0.0046

TDR

0.8900

0.8663

0.1894

0.5724

0.6198

0.4944

0.8565

60 * 60

IoU

0.4312

0.6603

0.1830

0.4580

0.2641

0.2795

0.5684

OD

0.0157

0.0191

0.0552

0.0347

0.0313

0.0494

0.0188

TDR

0.8788

0.8221

0.1028

0.4770

0.4310

0.2807

0.7209

  1. Bold indicates the best performance.. The dynamic trigger configurations includes: size (\(20 \times 20\), \(40 \times 40\), \(60 \times 60\)). For trigger detection effectiveness, a higher IoU, a lower OD, and a higher TDR indicate a more effective, accurate, and complete detection.