Skip to content

DATASET_CARD.md

Latest commit

 

History

History
50 lines (38 loc) · 1.77 KB

DATASET_CARD.md

File metadata and controls

50 lines (38 loc) · 1.77 KB
Raw

Dataset Card: Adversarial Plate Overlay Dataset

Reproducibility

The dataset generation process is deterministic when using the same:

  • Random seed (random_seed in config or --seed CLI option)
  • Input images (backgrounds and overlays)
  • Perturbation parameters
  • Software version

Reproducible Generation

# Generate identical datasets with same seed
advplate generate --seed 42 --config my_config.yaml

# Or specify in config file
# dataset:
#   random_seed: 42

All perturbations respect the random seed for consistent results across runs.

Purpose

This dataset generator is designed for research into adversarial robustness of Optical Character Recognition (OCR) and Automated License Plate Recognition (ALPR) models.
The goal is not to conceal license plates in real-world use, but to provide controlled, reproducible data for studying:

  • Model brittleness under adversarial overlays.
  • How adversarial training affects generalization.
  • Privacy-preserving methods for resisting surveillance capitalism and corporate misuse.

NOT for:

  • Evading law enforcement.
  • Facilitating illegal activity.
  • Deployment in production systems without ethical review.

Composition

  • Inputs: Background vehicle images, plate overlays (cropped PNGs).
  • Outputs: Composite images with adversarial perturbations + JSON metadata.

Use Cases

  • Benchmarking ALPR/OCR adversarial robustness.
  • Training defense models against adversarial perturbations.
  • Privacy/ethics research on surveillance systems.

Risks & Limitations

  • Generated examples are synthetic and may not reflect real-world conditions.
  • Misuse is possible if applied outside research contexts.

License

Released for academic and research use under MIT, subject to responsible disclosure and ethical use.