RAITAP evaluation/explainability use cases from the BA thesis.
Off-the-shelf COCO object detector (torchvision fasterrcnn_resnet50_fpn_v2)
run through RAITAP. mAP anchor (5000 imgs, mAP=0.4616, reproduces published
46.7) + per-box transparency on a 15-image subset.
Dermoscopy skin-lesion classification (HAM10000 / ISIC2018, 7 classes) with RAITAP 0.13.1. Two ResNet-50 models (adversarial eps1, vanilla). Transparency (SHAP / Grad-CAM / IG), metrics, robustness (PGD-L2), plus numerical + visual fidelity checks against the source libraries.
Certified L∞ robustness via auto_LiRPA (CROWN bound propagation) through RAITAP, on a small residual CNN (skip connections + BatchNorm) over a synthetic 4-colour set. Certifies 31/32 at ε=0.0015, plus a numerical fidelity check proving RAITAP's certified bounds are bit-identical to bare auto_LiRPA. The formal-verification companion to the Marabou MNIST use-case.
- Large assets (models, images, reports, attributions) tracked via Git LFS.
- See each subdir README for full detail.