Table 1 IoU results of generated saliency maps using seven XAI models during a poisoning attack with static stamping for detection effectiveness evaluation.

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

 

Sq.,Cn.,20

Sq.,Cn.,40

Sq.,Cn.,60

Sq.,Ct.,20

Sq.,Rd.,20

Cr.,Cn.,20

Cr.,Ct.,20

Cr.,Rd.,20

BP

0.3663

0.1994

0.1447

0.3961

0.3941

0.3906

0.5386

0.5663

Guided BP

0.5338

0.3061

0.2121

0.6682

0.6921

0.8573

0.7110

0.8185

GradCAM

0.0492

0.0378

0.1314

0.1980

0.3170

0.0607

0.1962

0.2030

Guided GradCAM

0.1424

0.0628

0.1232

0.5043

0.6963

0.5894

0.6274

0.7492

OS

0.2170

0.3343

0.5420

0.6485

0.3335

0.5139

0.6423

0.4130

Ablation

0.3261

0.0810

0.4358

0.1639

0.3159

0.6060

0.2457

0.4066

LIME

0.4989

0.5956

0.3711

0.7063

0.2844

0.7699

0.6437

0.0007

  1. Bold indicates the best, while underline indicates the worst. The static trigger configurations include shapes (square: Sq., circle: Cr.), positions (corner: Cn., center: Ct., random: Rd.), and size (\(20 \times 20\), \(40 \times 40\), \(60 \times 60\)). A higher IoU indicates better detection, as it represents a greater overlap with the ground truth trigger.
Back to article page