🔎 Edge Detection Attack Lab

---

1 . Project overview

This is a self-contained Python playground that simulates, attacks, and evaluates classical edge-detection pipelines (Sobel, Canny, Laplacian of Gaussian, Roberts Cross). It lets you

• preprocess and normalise a set of demo images;
• run a variety of targeted (algorithm-specific) and progressive (generic) adversarial attacks;
• compute effectiveness metrics (edge-density drop, contour-area reduction, fragmentation); and
• save side-by-side artefacts (clean / attacked / edge-maps) plus JSON scorecards for later analysis.

Everything is pure NumPy + OpenCV—no deep-learning toolkit required.

---

2 . Key components

Module	Purpose
targeted_attacks.py	Three attacks that focus on strong-edge pixels (blur, gradient reversal, contour disruption).
progressive_attacks.py	Pixel-noise escalation and generic edge/contour attacks executed over multiple “gentle → aggressive” rounds.
canny_targeted_attacks.py	Five attacks that exploit specific Canny stages (hysteresis noise, gradient smoothing, non-max-suppression confusion, connectivity gaps, multi-scale perturbations) and sweep three Canny hyper-parameter sets.
simulation_engine.py	Object-oriented façade (EdgeAttackSimulationEngine) with > 40 tunable parameters – useful for interactive notebooks or web front-ends.

Each script deposits PNG artefacts plus a *_results.json manifest in its own output folder (targeted_attacks/, progressive_attacks/, …).

---

3 . Installation

# 1. Clone and cd
git clone https://github.com/<you>/edge-attack-lab.git
cd edge-attack-lab

# 2. Create environment (≈ Python 3.9+)
python -m venv .venv
source .venv/bin/activate                # Linux / macOS
# .venv\Scripts\activate.bat             # Windows

# 3. Install runtime deps
pip install -r requirements.txt
# requirements.txt
# ├─ numpy
# └─ opencv-python

Tip: add opencv-contrib-python if you later need advanced CV functions.

---

4 . Running the demos

# 1. Generic Sobel / contour / gradient demo
python targeted_attacks.py

# 2. Six-stage progressive escalation
python progressive_attacks.py

# 3. Canny-specific research sweep
python canny_targeted_attacks.py

# 4. Custom batch with OO engine
python simulation_engine.py

Each run ends with a concise success tally:

Results saved to 'progressive_attacks/'
Successful attacks: 11/18

and drops artefacts such as:

progressive_attacks/
│  street_scene-sobel-gentle_pixels-clean.png
│  street_scene-sobel-gentle_pixels-pert.png
│  street_scene-sobel-gentle_pixels-edges-clean.png
│  street_scene-sobel-gentle_pixels-edges-pert.png
└─ metadata.json

---

5 . Understanding the metrics

Metric	Meaning	Success threshold
edge_density_reduction	Fractional drop in total edge pixels.	> 15 %
contour_area_reduction	Drop in largest connected edge region.	> 20 %
fragmentation_increase	(Canny-only) rise in contour count.	> 50 %
attack_success	True if any threshold is crossed.	–

---

6 . Extending the lab

1. Add a new attack Implement a function that takes gray_image: np.ndarray and returns a perturbed image of identical shape. Register it inside the relevant attack_functions dictionary.

2. Plug in your own images Drop them into source_images/ and extend the IMAGE_SOURCES mapping.

3. Integrate with notebooks / web UI Import EdgeAttackSimulationEngine, tweak SimulationConfig, and call execute_simulation_on_image(...).

---

7 . Roadmap

• 🔬 Add patch-wise learned perturbations via gradient-free optimisation.
• 🖥 Streamline a Streamlit or Next.js front-end for interactive exploration.
• 📊 Auto-generate HTML reports comparing detectors, attacks, and thresholds.

---

8 . Contributing & licence

Pull requests and issue reports are welcome!