Update test/adversarial_test.py

test/adversarial_test.py

@@ -1,67 +1,32 @@
-# !pip install adversarial-robustness-toolbox requests
+# test/adversarial_test.py
 
-import requests
-import numpy as np
-from art.estimators.classification import BlackBoxClassifier
-from art.attacks.evasion import TextAttackExtraction, WordSwapEmbedding
+from art.estimators.language import HuggingFacePipeline
+from art.attacks.text import TextAttackExtraction, TextFooler
+from transformers import pipeline
+import os
 
-# -------------------------
-# ۱) آدرس API اسپیس شما
-# -------------------------
-API_URL = "https://code5ecure.hf.space/run/predict"
+# Load your model pipeline from app.py
+# Replace this with your actual function if you wrapped it differently
+from app import generator  
 
-def query_space(prompt: str) -> str:
-    """
-    تابعی برای ارسال prompt به اسپیس و دریافت خروجی
-    """
-    response = requests.post(API_URL, json={"data": [prompt]})
-    try:
-        return response.json()["data"][0]
-    except Exception:
-        return "Error or unexpected output"
+# Wrap the HF pipeline with ART
+art_model = HuggingFacePipeline(pipeline=generator)
 
-# -------------------------
-# ۲) Wrapper برای ART BlackBoxClassifier
-# -------------------------
-def predict_fn(texts):
-    """
-    برای تست ساده: هر متن رو به classifier می‌دهیم و خروجی 1 کلاس برمی‌گرده
-    """
-    preds = []
-    for t in texts:
-        out = query_space(t)
-        # چون TextGeneration داریم، کلاس واقعی نداریم
-        preds.append([1.0])
-    return np.array(preds)
+# Example prompt to test
+prompt = "Tell me a positive thing about flowers in Africa."
 
-classifier = BlackBoxClassifier(
-    predict=predict_fn,
-    input_shape=(1,),
-    nb_classes=1,
-    clip_values=(0, 1),
-)
+# === Attack 1: TextAttackExtraction ===
+textattack = TextAttackExtraction(estimator=art_model)
+adv_examples_extraction = textattack.generate([prompt])
 
-# -------------------------
-# ۳) تعریف حملات
-# -------------------------
-attacks = {
-    "TextAttackExtraction": TextAttackExtraction(classifier),
-    "WordSwapEmbedding": WordSwapEmbedding(classifier)
-}
+print("TextAttackExtraction Adversarial Examples:")
+for ex in adv_examples_extraction:
+    print(ex)
 
-# -------------------------
-# ۴) متن تست اولیه
-# -------------------------
-texts = ["Write me a nice poem about flowers."]
+# === Attack 2: TextFooler ===
+textfooler = TextFooler(estimator=art_model)
+adv_examples_fooler = textfooler.generate([prompt])
 
-# -------------------------
-# ۵) اجرای حملات و چاپ نتایج
-# -------------------------
-for name, attack in attacks.items():
-    adv_texts = attack.generate(texts)
-    print("===================================")
-    print(f"🟢 حمله: {name}")
-    print("متن اصلی:", texts[0])
-    print("متن مخرب:", adv_texts[0])
-    print("پاسخ مدل به متن اصلی:", query_space(texts[0]))
-    print("پاسخ مدل به متن مخرب:", query_space(adv_texts[0]))
+print("\nTextFooler Adversarial Examples:")
+for ex in adv_examples_fooler:
+    print(ex)