BETTER-NMA 2.0.2

NMA: Dendrogram-based model analysis, white-box testing, and adversarial detection

---

NMA – Near Misses Analysis

NMA (Near Misses Analysis) is a Python package for analyzing machine learning models through dendrogram-based hierarchical clustering, white-box testing, and adversarial attack detection.

It provides visualization, explanation, and diagnostic tools to help developers and researchers understand their models’ decision boundaries, identify vulnerabilities, and detect adversarial inputs.

---

✨ Features

• 📊 Dendrogram construction & visualization

  • Build hierarchical trees from model predictions.
  • Plot full dendrograms or sub-dendrograms for specific labels.

• 🌳 Concept Tree Generation & Plotting

  • Extract semantic WordNet concept hierarchies from dendrograms.
  • Generic fallback support: custom dictionaries or auto WordNet search.
  • Publication-ready colored concept tree visualization.

• 🧪 White-box testing

  • Identify problematic training samples likely to cause misclassification.
  • Run structured analysis across source/target label pairs.

• 🛡 Adversarial attack detection

  • Train a logistic regression adversarial detector.
  • Detect adversarial images and compute adversarial scores.

• 🔎 Model querying & explanations

  • Query images for predictions with hierarchical context.
  • Generate verbal explanations of model predictions.

• 🧩 Cluster analysis tools

  • Find lowest common ancestors (LCA) in the dendrogram.
  • Rename clusters for more meaningful interpretation.

---

📦 Installation

pip install BETTER_NMA

---

🚀 Quickstart

from BETTER_NMA import NMA
import numpy as np

# Example data (replace with your dataset/model)
x_train = np.random.rand(100, 32, 32, 3)
y_train = np.random.randint(0, 2, size=100)
labels = ["cat", "dog"]

# Your pre-trained model (e.g., Keras, PyTorch wrapper with predict)
model = my_model  

# Initialize NMA
nma = NMA(
    x_train=x_train,
    y_train=y_train,
    labels=labels,
    model=model,
    explanation_method="similarity", 
    save_connections=True
)

# Plot dendrogram
nma.plot(title="Model Decision Hierarchy")

# Run white-box testing
issues = nma.white_box_testing(["cat"], ["dog"], analyze_results=True)

# Train adversarial detector
nma.train_adversarial_detector(authentic_images, adversarial_images)

# Detect if a new image is adversarial
result = nma.detect_attack(test_image)

# Get verbal explanation of an image
explanation = nma.verbal_explanation(test_image)
print(explanation)

# Build a semantic Concept Tree using WordNet
# Defaults to built-in ImageNet mapping, or fallback text-search if no mapping match
concept_tree = nma.build_concept_tree()

# Plot the resolved Concept Tree
nma.plot_concept_tree(concept_tree, title="Decision Concept Tree")

---

📚 API Overview

Dendrogram & Visualization

• plot(sub_labels=None, ...) – plot full or partial dendrogram.
• plot_sub_dendrogram(sub_labels, ...) – zoom into specific classes.
• build_concept_tree(custom_mapping=None, allowed_labels=None) – build resolved Concept Tree.
• plot_concept_tree(concept_tree, title=...) – plot colored multi-branch concept hierarchy.

White-box Testing

• white_box_testing(source_labels, target_labels, ...) – find problematic images.
• get_white_box_analysis(source_labels, target_labels, ...) – detailed analysis.

Adversarial Detection

• train_adversarial_detector(authentic_images, attacked_images) – train detector.
• detect_attack(image, plot_result=False) – detect adversarial samples using logistic regression detector.
• detect_attack_by_threshold(image, threshold=0.35, ...) – detect adversarial samples using a specific LCA score threshold.
• adversarial_score(image, top_k=5) – compute adversarial score.

Query & Explanation

• query_image(image, top_k=5) – get predictions & explanation.
• verbal_explanation(image) – generate natural language explanation.

Cluster Analysis

• find_lca(label1, label2) – lowest common ancestor.
• change_cluster_name(cluster_id, new_name) – rename clusters.

---

🌳 Concept Tree Guide

The Concept Tree feature maps your model's classification hierarchy into a human-understandable WordNet semantic tree by resolving intermediate clusters.

Custom database / label-to-synset mappings

If you are working with a custom dataset or custom labels, you can pass a custom dictionary mapping class labels to WordNet synset IDs:

# Custom mapping: key is label, value is ImageNet/WordNet synset ID (n + 8 digits)
custom_map = {
    "cat": "n02121808",
    "dog": "n02084071",
    "apple": "n07739125",
}

# Build the tree using your custom mapping
concept_tree = nma.build_concept_tree(custom_mapping=custom_map)
nma.plot_concept_tree(concept_tree)

If no mapping is provided, NMA will use the default 41-class ImageNet mapping. If a class label doesn't match any mapping, NMA automatically falls back to searching WordNet directly using the label name or sub-words.

---

🛠 Requirements

• Python ≥ 3.8
• NumPy, Pandas, Matplotlib, Scikit-learn, SciPy
• NLTK (requires wordnet corpus downloaded: nltk.download('wordnet'))
• (Optional) PyTorch / TensorFlow for model support

---

📖 Use Cases

• Research – interpret model predictions via hierarchical clustering.
• Robustness testing – identify adversarial vulnerabilities.
• Explainability – provide visual + verbal explanations.
• Debugging – detect mislabeled or problematic training samples.

---

📜 License

MIT License – free to use and modify.

---