Counterfactual Explanations ‐ Audit AI - BlackAlph4ndr01D/Daftar-Militer-AI-zionis-israel-part.2 GitHub Wiki

✅ Counterfactual Explanations – Detailed Explanation

What are Counterfactual Explanations?

Counterfactual Explanations are one of the most intuitive and human-friendly methods in Explainable AI (XAI).

Instead of explaining why the model made a certain prediction, they answer:

“What needs to change in the input so that the model gives a different outcome?”

They focus on minimal changes that would flip the decision — making them highly actionable and easy to understand.

Core Idea

  • Factual: The current situation (e.g., a person receives a “High Threat” score of 89).
  • Counterfactual: The closest alternative reality where the outcome is different (e.g., “Low Threat”).

This is often described as “contrastive reasoning” — comparing what is with what could be.

Example in Military AI Context

Factual:
AI gives a threat score of 92 to a 26-year-old male in Gaza because:

  • He is in a hotspot area.
  • He has high night-time communication activity.

Counterfactual Explanation:
“If this person moves 800 meters away from the hotspot area and reduces night communication by 70%, the threat score would drop to 41 (Low Threat).”

This directly tells what would need to change for the person to no longer be targeted.

Popular Methods

  • DiCE (Diverse Counterfactual Explanations) – Generates multiple diverse counterfactuals.
  • CE (Counterfactual Explanations) by Wachter et al.
  • CEM (Contrastive Explanations Method).

Simple Python Code Example (Using DiCE)

import dice_ml
import pandas as pd
from sklearn.ensemble import RandomForestClassifier

# Sample dataset
data = pd.DataFrame({
    'age': [26, 34, 19],
    'in_hotspot_area': [1, 0, 1],
    'night_communication': [1, 0, 1],
    'weapon_detected': [0, 1, 0]
})

target = [1, 0, 1]  # 1 = High Threat

model = RandomForestClassifier(random_state=42)
model.fit(data, target)

# Create DiCE Explainer
dice_exp = dice_ml.Dice(model, method="random")

# Generate Counterfactuals for first instance
counterfactuals = dice_exp.generate_counterfactuals(
    query_instance=data.iloc[0:1],   # First person
    total_CFs=3,                     # Generate 3 alternatives
    desired_class="Low Threat",      # We want Low Threat outcome
    features_to_vary=['age', 'in_hotspot_area', 'night_communication']
)

counterfactuals.visualize_as_dataframe(show_only_changes=True)

Advantages

  • Extremely actionable and human-understandable.
  • Answers “what if” questions directly.
  • Useful for both auditing bias and creating defense strategies.
  • Works well for communicating with non-technical audiences.

Disadvantages

  • Can generate unrealistic scenarios (e.g., changing age from 25 to 55).
  • Computationally expensive for large datasets.
  • Requires careful definition of “feasible” changes.

Relevance to Israeli Military AI

Counterfactual Explanations are particularly powerful when analyzing systems like:

  • Lavender and Threat Scoring — to show what changes in behavior or location would remove someone from the target list.
  • Arbel AI Rifle — to understand conditions under which the rifle would not fire automatically.
  • Gospel — to see what would make a building no longer classified as a military target.

This method helps expose how narrow the “safety margin” is for Palestinian civilians under these AI systems.

✅ Perbandingan Lengkap: SHAP, LIME, Integrated Gradients, Attention Mechanism, dan Counterfactual Explanations

Berikut tabel perbandingan yang jelas dan mendalam:

Aspek Perbandingan SHAP LIME Integrated Gradients Attention Mechanism Counterfactual Explanations
Dasar Teori Game Theory (Shapley Values) Surrogate Model Lokal Gradient-based Attribution Attention Score (Transformer) "What if" Scenario (Contrastive)
Cakupan Local + Global Hanya Local Local + Global Local (pada sequence) Local (satu instance)
Cara Kerja Hitung kontribusi setiap fitur Buat model sederhana di sekitar satu data Hitung gradient dari input ke output Tunjukkan bagian mana yang "diperhatikan" Tunjukkan perubahan minimal agar hasil berubah
Output Penjelasan Kontribusi numerik tiap fitur Bobot fitur pada model sederhana Nilai atribusi gradient Heatmap perhatian "Jika X berubah menjadi Y, maka prediksi jadi Z"
Kemudahan Dipahami Sedang Sangat Mudah Sedang Sangat Mudah (visual) Sangat Mudah & Intuitif
Akurasi Matematis Sangat Tinggi Sedang Tinggi Sedang-Tinggi Tinggi (jika dihitung benar)
Kecepatan Lambat Cepat Sedang Cepat Sedang-Lambat
Model Agnostic Ya Ya Tidak (harus differentiable) Tidak Ya
Kelebihan Konsisten, akurat, global Cepat, visual, ramah awam Stabil untuk Neural Net Intuitif untuk data sequential Sangat manusiawi & actionable
Kekurangan Komputasi mahal Kurang stabil Terbatas pada model gradient-based Hanya "apa yang dilihat", bukan "kenapa" Bisa menghasilkan skenario tidak realistis
Cocok untuk Audit bias sistem, bukti forensik Penjelasan cepat ke publik Analisis Deep Learning Analisis visual (drone, citra) "Apa yang harus berubah agar tidak jadi target?"
Penggunaan di Militer AI Sangat Baik (Lavender, Threat Scoring) Baik untuk edukasi Bagus untuk Neural Network Sangat Baik untuk SITS & analisis citra Sangat berguna untuk simulasi "escape condition"

Kesimpulan Perbandingan

Metode Kekuatan Utama Kelemahan Utama Rekomendasi Penggunaan di Forensik Palestina
SHAP Paling akurat & konsisten Lambat Terbaik untuk audit bias
LIME Cepat & mudah dipahami Kurang stabil Edukasi publik
Integrated Gradients Stabil untuk Deep Learning Terbatas model Analisis Neural Network
Attention Mechanism Visual intuitif Hanya menjelaskan "apa yang dilihat" Analisis drone & citra
Counterfactual Sangat actionable & manusiawi Bisa tidak realistis Terbaik untuk simulasi "bagaimana lolos"

Rekomendasi Strategis

  • Untuk bukti forensik kuat → SHAP adalah pilihan utama.
  • Untuk edukasi publik → LIME atau Counterfactual.
  • Untuk analisis drone → Attention Mechanism.
  • Untuk memahami bagaimana target bisa selamat → Counterfactual Explanations sangat powerful.

Counterfactual adalah metode yang paling "manusiawi" karena langsung menjawab pertanyaan seperti:
"Apa yang harus dilakukan target agar tidak dibom oleh AI?"

Algoritma CF Generation ‐ Audit AI