All projects
PerceptionResearch

Traffic-Sign Attack Generator

Physical adversarial shadows and occlusions for stress-testing traffic-sign recognition.

Stop sign with a shadow attack

Overview

A configurable generator that applies physically plausible attacks to any traffic-sign dataset described by a CSV manifest. It covers triangular shadows inside the sign region and occlusions (edge patch, center sticker, diagonal band), with optional rotation, brightness, contrast and blur.

Every output is written alongside a manifest recording the exact attack parameters and ROI metadata, so robustness experiments are fully reproducible.

The generator powered a group project on adversarial robustness. An AlexNet classifier for the 43 GTSRB sign classes was retrained on mixes of clean and attacked images. It was compared with a defence pipeline that estimates the probability of an attack, denoises the image, and blends the corrected and original images in proportion to that probability before classifying.

Highlights

  • 01Shape-aware masks: circle, triangle, diamond, octagon and box.
  • 02Dataset-agnostic CSV manifest with optional ROIs.
  • 03Reproducible output manifests for every attacked image.
  • 04Used for adversarial training and for an attack-detection and denoising defence that can be fine-tuned end to end.

Gallery

Edge-patch occlusion.
Shadow on a yield sign.
Sticker-style occlusion.

Command menu

Jump to a section or project