updated to translate final message to original language detected

app.py

@@ -8,7 +8,7 @@ from deep_translator import GoogleTranslator
 import openai
 import os
 
-# Set up your OpenAI API key
+# Set your OpenAI API key
 openai.api_key = os.getenv("OPENAI_API_KEY")
 
 # Translator instance
@@ -120,7 +120,6 @@ def query_llm_for_classification(raw_message: str) -> dict:
         raw_reply = response["choices"][0]["message"]["content"].strip()
 
         import json
-        # Expect something like {"label": "...", "reason": "..."}
         llm_result = json.loads(raw_reply)
         if "label" not in llm_result or "reason" not in llm_result:
             return {"label": "Unknown", "reason": f"Unexpected format: {raw_reply}"}
@@ -170,19 +169,31 @@ def query_llm_for_explanation(
     llm_reason: str,
     found_smishing: list,
     found_other_scam: list,
-    found_urls: list
+    found_urls: list,
+    detected_lang: str
 ) -> str:
     """
-    Second LLM call: provides a holistic explanation of the final classification.
-    We include the local classification info, the LLM’s own classification, and
-    relevant details (keywords, URLs).
+    Second LLM call: provides a holistic explanation of the final classification
+    in the same language as detected_lang (English or Spanish).
     """
-    system_prompt = (
-        "You are a cybersecurity expert providing a final explanation to the user. "
-        "Combine the local classification, the LLM classification, and the final label "
-        "into one concise explanation. Do not reveal internal code or raw JSON; just give "
-        "a short, user-friendly explanation. End with a final statement of the final label."
-    )
+    # Decide the language for final explanation
+    if detected_lang == "es":
+        # Spanish
+        system_prompt = (
+            "Eres un experto en ciberseguridad. Proporciona una explicación final al usuario en español. "
+            "Combina la clasificación local, la clasificación LLM y la etiqueta final en una sola explicación breve. "
+            "No reveles el código interno ni el JSON bruto; simplemente da una breve explicación fácil de entender. "
+            "Termina con la etiqueta final. "
+        )
+    else:
+        # Default to English
+        system_prompt = (
+            "You are a cybersecurity expert providing a final explanation to the user in English. "
+            "Combine the local classification, the LLM classification, and the final label "
+            "into one concise explanation. Do not reveal internal code or raw JSON. "
+            "End with a final statement of the final label."
+        )
+
     user_context = f"""
 User Message:
 {text}
@@ -195,7 +206,7 @@ Suspicious SMiShing Keywords => {found_smishing}
 Suspicious Other Scam Keywords => {found_other_scam}
 URLs => {found_urls}
 """
-    # The LLM can combine these facts into a short paragraph.
+
     try:
         response = openai.ChatCompletion.create(
             model="gpt-3.5-turbo",
@@ -215,7 +226,7 @@ def smishing_detector(input_type, text, image):
     Main detection function combining text (if 'Text') & OCR (if 'Screenshot'),
     plus two LLM calls: 
       1) classification to adjust final probabilities,
-      2) a final explanation summarizing the outcome.
+      2) a final explanation summarizing the outcome in the detected language.
     """
     if input_type == "Text":
         combined_text = text.strip() if text else ""
@@ -275,7 +286,7 @@ def smishing_detector(input_type, text, image):
     found_smishing = [kw for kw in smishing_keys if kw in lower_text]
     found_other_scam = [kw for kw in scam_keys if kw in lower_text]
 
-    # 6. Final LLM explanation
+    # 6. Final LLM explanation (in detected_lang)
     final_explanation = query_llm_for_explanation(
         text=combined_text,
         final_label=final_label,
@@ -286,7 +297,8 @@ def smishing_detector(input_type, text, image):
         llm_reason=llm_reason,
         found_smishing=found_smishing,
         found_other_scam=found_other_scam,
-        found_urls=found_urls
+        found_urls=found_urls,
+        detected_lang=detected_lang
     )
 
     return {
@@ -321,7 +333,7 @@ def toggle_inputs(choice):
         return gr.update(visible=False), gr.update(visible=True)
 
 with gr.Blocks() as demo:
-    gr.Markdown("## SMiShing & Scam Detector with LLM-Enhanced Logic")
+    gr.Markdown("## SMiShing & Scam Detector with LLM-Enhanced Logic (Multilingual Explanation)")
     
     with gr.Row():
         input_type = gr.Radio(
@@ -363,7 +375,6 @@ with gr.Blocks() as demo:
     )
 
 if __name__ == "__main__":
-    # Warn if openai.api_key not set
     if not openai.api_key:
         print("WARNING: OPENAI_API_KEY not set. LLM calls will fail or be skipped.")
     demo.launch()

app.py.bestoftues

@@ -5,6 +5,11 @@ from transformers import pipeline
 import re
 from langdetect import detect
 from deep_translator import GoogleTranslator
+import openai
+import os
+
+# Set up your OpenAI API key
+openai.api_key = os.getenv("OPENAI_API_KEY")
 
 # Translator instance
 translator = GoogleTranslator(source="auto", target="es")
@@ -23,7 +28,7 @@ CANDIDATE_LABELS = ["SMiShing", "Other Scam", "Legitimate"]
 
 def get_keywords_by_language(text: str):
     """
-    Detect language using `langdetect` and translate keywords if needed.
+    Detect language using langdetect and translate keywords if needed.
     """
     snippet = text[:200]
     try:
@@ -88,9 +93,129 @@ def boost_probabilities(probabilities: dict, text: str):
         "detected_lang": detected_lang
     }
 
+def query_llm_for_classification(raw_message: str) -> dict:
+    """
+    First LLM call: asks for a classification (SMiShing, Other Scam, or Legitimate)
+    acting as a cybersecurity expert. Returns label and short reason.
+    """
+    if not raw_message.strip():
+        return {"label": "Unknown", "reason": "No message provided to the LLM."}
+
+    system_prompt = (
+        "You are a cybersecurity expert. You will classify the user's message "
+        "as one of: SMiShing, Other Scam, or Legitimate. Provide a short reason. "
+        "Return only JSON with keys: label, reason."
+    )
+    user_prompt = f"Message: {raw_message}\nClassify it as SMiShing, Other Scam, or Legitimate."
+
+    try:
+        response = openai.ChatCompletion.create(
+            model="gpt-3.5-turbo",
+            messages=[
+                {"role": "system", "content": system_prompt},
+                {"role": "user", "content": user_prompt}
+            ],
+            temperature=0.2
+        )
+        raw_reply = response["choices"][0]["message"]["content"].strip()
+
+        import json
+        # Expect something like {"label": "...", "reason": "..."}
+        llm_result = json.loads(raw_reply)
+        if "label" not in llm_result or "reason" not in llm_result:
+            return {"label": "Unknown", "reason": f"Unexpected format: {raw_reply}"}
+
+        return llm_result
+
+    except Exception as e:
+        return {"label": "Unknown", "reason": f"LLM error: {e}"}
+
+def incorporate_llm_label(boosted: dict, llm_label: str) -> dict:
+    """
+    Adjust the final probabilities based on the LLM's classification.
+    If LLM says SMiShing, add +0.2 to SMiShing, etc. Then clamp & re-normalize.
+    """
+    if llm_label == "SMiShing":
+        boosted["SMiShing"] += 0.2
+    elif llm_label == "Other Scam":
+        boosted["Other Scam"] += 0.2
+    elif llm_label == "Legitimate":
+        boosted["Legitimate"] += 0.2
+    # else "Unknown" => do nothing
+
+    # clamp
+    for k in boosted:
+        if boosted[k] < 0:
+            boosted[k] = 0.0
+
+    total = sum(boosted.values())
+    if total > 0:
+        for k in boosted:
+            boosted[k] /= total
+    else:
+        # fallback
+        boosted["Legitimate"] = 1.0
+        boosted["SMiShing"] = 0.0
+        boosted["Other Scam"] = 0.0
+
+    return boosted
+
+def query_llm_for_explanation(
+    text: str,
+    final_label: str,
+    final_conf: float,
+    local_label: str,
+    local_conf: float,
+    llm_label: str,
+    llm_reason: str,
+    found_smishing: list,
+    found_other_scam: list,
+    found_urls: list
+) -> str:
+    """
+    Second LLM call: provides a holistic explanation of the final classification.
+    We include the local classification info, the LLM’s own classification, and
+    relevant details (keywords, URLs).
+    """
+    system_prompt = (
+        "You are a cybersecurity expert providing a final explanation to the user. "
+        "Combine the local classification, the LLM classification, and the final label "
+        "into one concise explanation. Do not reveal internal code or raw JSON; just give "
+        "a short, user-friendly explanation. End with a final statement of the final label."
+    )
+    user_context = f"""
+User Message:
+{text}
+
+Local Classification => Label: {local_label}, Confidence: {local_conf}
+LLM Classification => Label: {llm_label}, Reason: {llm_reason}
+Final Overall Label => {final_label} (confidence {final_conf})
+
+Suspicious SMiShing Keywords => {found_smishing}
+Suspicious Other Scam Keywords => {found_other_scam}
+URLs => {found_urls}
+"""
+    # The LLM can combine these facts into a short paragraph.
+    try:
+        response = openai.ChatCompletion.create(
+            model="gpt-3.5-turbo",
+            messages=[
+                {"role": "system", "content": system_prompt},
+                {"role": "user", "content": user_context}
+            ],
+            temperature=0.2
+        )
+        final_explanation = response["choices"][0]["message"]["content"].strip()
+        return final_explanation
+    except Exception as e:
+        return f"Could not generate final explanation due to error: {e}"
+
 def smishing_detector(input_type, text, image):
     """
-    Main detection function combining text (if 'Text') and OCR (if 'Screenshot').
+    Main detection function combining text (if 'Text') & OCR (if 'Screenshot'),
+    plus two LLM calls: 
+      1) classification to adjust final probabilities,
+      2) a final explanation summarizing the outcome.
     """
     if input_type == "Text":
         combined_text = text.strip() if text else ""
@@ -105,31 +230,44 @@ def smishing_detector(input_type, text, image):
             "label": "No text provided",
             "confidence": 0.0,
             "keywords_found": [],
-            "urls_found": []
+            "urls_found": [],
+            "llm_label": "Unknown",
+            "llm_reason": "No text to analyze",
+            "final_explanation": "No text provided"
         }
 
-    result = classifier(
+    # 1. Local zero-shot classification
+    local_result = classifier(
         sequences=combined_text,
         candidate_labels=CANDIDATE_LABELS,
         hypothesis_template="This message is {}."
     )
-    original_probs = {k: float(v) for k, v in zip(result["labels"], result["scores"])}
+    original_probs = {k: float(v) for k, v in zip(local_result["labels"], local_result["scores"])}
 
+    # 2. Basic boosting from keywords & URLs
     boosted = boost_probabilities(original_probs, combined_text)
+    detected_lang = boosted.pop("detected_lang", "en")
+
+    # Convert to float only
+    for k in boosted:
+        boosted[k] = float(boosted[k])
+
+    local_label = max(boosted, key=boosted.get)
+    local_conf = round(boosted[local_label], 3)
 
-    # Patched snippet begins
-    # 1. Extract language first, preserving it
-    detected_lang = boosted.get("detected_lang", "en")
-    # 2. Remove it so only numeric keys remain
-    boosted.pop("detected_lang", None)
-    # 3. Convert numeric values to float
-    for k, v in boosted.items():
-        boosted[k] = float(v)
-    # Patched snippet ends
+    # 3. LLM Classification
+    llm_classification = query_llm_for_classification(combined_text)
+    llm_label = llm_classification.get("label", "Unknown")
+    llm_reason = llm_classification.get("reason", "No reason provided")
 
+    # 4. Incorporate LLM’s label into final probabilities
+    boosted = incorporate_llm_label(boosted, llm_label)
+
+    # Now we have updated probabilities
     final_label = max(boosted, key=boosted.get)
     final_confidence = round(boosted[final_label], 3)
 
+    # 5. Gather found keywords & URLs
     lower_text = combined_text.lower()
     smishing_keys, scam_keys, _ = get_keywords_by_language(combined_text)
 
@@ -137,16 +275,34 @@ def smishing_detector(input_type, text, image):
     found_smishing = [kw for kw in smishing_keys if kw in lower_text]
     found_other_scam = [kw for kw in scam_keys if kw in lower_text]
 
+    # 6. Final LLM explanation
+    final_explanation = query_llm_for_explanation(
+        text=combined_text,
+        final_label=final_label,
+        final_conf=final_confidence,
+        local_label=local_label,
+        local_conf=local_conf,
+        llm_label=llm_label,
+        llm_reason=llm_reason,
+        found_smishing=found_smishing,
+        found_other_scam=found_other_scam,
+        found_urls=found_urls
+    )
+
     return {
         "detected_language": detected_lang,
         "text_used_for_classification": combined_text,
         "original_probabilities": {k: round(v, 3) for k, v in original_probs.items()},
-        "boosted_probabilities": {k: round(v, 3) for k, v in boosted.items()},
+        "boosted_probabilities_before_llm": {local_label: local_conf},
+        "llm_label": llm_label,
+        "llm_reason": llm_reason,
+        "boosted_probabilities_after_llm": {k: round(v, 3) for k, v in boosted.items()},
         "label": final_label,
         "confidence": final_confidence,
         "smishing_keywords_found": found_smishing,
         "other_scam_keywords_found": found_other_scam,
         "urls_found": found_urls,
+        "final_explanation": final_explanation,
     }
 
 #
@@ -165,7 +321,7 @@ def toggle_inputs(choice):
         return gr.update(visible=False), gr.update(visible=True)
 
 with gr.Blocks() as demo:
-    gr.Markdown("## SMiShing & Scam Detector (Choose Text or Screenshot)")
+    gr.Markdown("## SMiShing & Scam Detector with LLM-Enhanced Logic")
     
     with gr.Row():
         input_type = gr.Radio(
@@ -207,4 +363,7 @@ with gr.Blocks() as demo:
     )
 
 if __name__ == "__main__":
+    # Warn if openai.api_key not set
+    if not openai.api_key:
+        print("WARNING: OPENAI_API_KEY not set. LLM calls will fail or be skipped.")
     demo.launch()