Research Reproduction

SPAR

Street-Legal Physical Adversarial Rim — arXiv:2604.02457

Reproduction of the SPAR attack: an adversarial border around a license plate that fools ALPR systems (fast-alpr: YOLO-v9 detection + cct-xs OCR). The paper's results: 60% reduction in correct reads, 18% impersonation rate, under $100 total cost.

⚠️ Legal note (Georgia): OCGA § 40-2-41 prohibits apparatus that obstructs or hinders the clear display and legibility of a license plate. The rim design sits behind the plate and covers no characters, which is fightable but untested. Do not deploy on a public road without legal review. Full analysis on the Legal page.

Start here

The Plan →
The full 7-step reproduction pipeline: photos → label → train → print → mount → evaluate.
Web Labeler →
Run the labeling party, controls reference, and the improvement roadmap.
Legal →
Georgia law notes, case law, and where SPAR sits.

Layout

web_labeler/
Corner-labeling web GUI — the labeling-party tool
train_patch.py
Training loop (kornia homography warp + det/OCR/TV losses)
losses.py
Loss functions
data_loader.py
Data loading + train/val split
generate_synthetic_data.py
Synthetic plate-image generator
evaluate.py
Digital evaluation + metrics
render_patch.py
Patch tensor → 300 DPI printable PNG
label_corners.py
Legacy OpenCV CLI labeler (web_labeler preferred)
check_labels.py
Label validation + verification collage
data/
raw/ (training images) · labels/ (corner JSON) · eval/
patches/
Trained patch weights + printable renders
results/
Eval CSV + summary

Data status (2026-08-15)

References