app.py

import gradio as gr
from requests import head
from transformer_vectorizer import TransformerVectorizer
from sklearn.feature_extraction.text import TfidfVectorizer
import numpy as np 
from concrete.ml.deployment import FHEModelClient
import numpy
import os
from pathlib import Path
import requests
import json
import base64
import subprocess
import shutil
import time
import easyocr
import PyPDF2
import os

reader = easyocr.Reader(['en']) 
# This repository's directory
REPO_DIR = Path(__file__).parent

subprocess.Popen(["uvicorn", "server:app"], cwd=REPO_DIR)

# Wait 5 sec for the server to start
time.sleep(5)

# Encrypted data limit for the browser to display
# (encrypted data is too large to display in the browser)
ENCRYPTED_DATA_BROWSER_LIMIT = 500
N_USER_KEY_STORED = 20
model_names=['financial_rating','legal_rating']


FHE_MODEL_PATH = "deployment/financial_rating"
FHE_LEGAL_PATH = "deployment/legal_rating"
#FHE_LEGAL_PATH="deployment/legal_rating"

print("Loading the transformer model...")

# Initialize the transformer vectorizer
transformer_vectorizer = TransformerVectorizer()
vectorizer = TfidfVectorizer()
def process_input(input_type, user_input, uploaded_file):
    print('ooooocr')
    if input_type == "File Upload" and uploaded_file is not None:
        # 读取上传的文件
        with open(uploaded_file.name, "rb") as f:
            image = f.read()
        results = reader.readtext(image)
        # 提取识别的文本
        extracted_text = ' '.join([text[1] for text in results])
        print("提取的文本:")
        print(extracted_text)
        return extracted_text
    elif input_type == "Text Input":
        return user_input
'''        
def process_input(input_type, user_input, uploaded_file):
    if input_type == "File Upload" and uploaded_file is not None:
        file_ext = os.path.splitext(uploaded_file.name)[1].lower()
        extracted_text = ""
        
        if file_ext in ['.jpg', '.jpeg', '.png']:
            # 处理图片文件
            results = reader.readtext(uploaded_file.name)
            extracted_text = ' '.join([text[1] for text in results])
            print("从图片提取的文本:")
            print(extracted_text)
        
        elif file_ext == '.txt':
            # 处理TXT文件
            with open(uploaded_file.name, 'r', encoding='utf-8') as f:
                extracted_text = f.read()
            print("从TXT文件提取的文本:")
            print(extracted_text)
        
        elif file_ext == '.pdf':
            # 处理PDF文件
            with open(uploaded_file.name, 'rb') as f:
                reader_pdf = PyPDF2.PdfReader(f)
                for page_num in range(len(reader_pdf.pages)):
                    page = reader_pdf.pages[page_num]
                    extracted_text += page.extract_text() + "\n"
            print("从PDF文件提取的文本:")
            print(extracted_text)
        
        else:
            return "不支持的文件类型。请上传 .jpg, .jpeg, .png, .txt 或 .pdf 文件。"
        
        return extracted_text

    elif input_type == "Text Input":
        return user_input
    else:
        return "无效的输入类型或未上传文件。"
'''
        
def toggle_visibility(input_type):
    user_input_visible = input_type == "Text Input"
    file_upload_visible = input_type == "File Upload"
    return gr.update(visible=user_input_visible), gr.update(visible=file_upload_visible)


def clean_tmp_directory():
    # Allow 20 user keys to be stored.
    # Once that limitation is reached, deleted the oldest.
    path_sub_directories = sorted([f for f in Path(".fhe_keys/").iterdir() if f.is_dir()], key=os.path.getmtime)

    user_ids = []
    if len(path_sub_directories) > N_USER_KEY_STORED:
        n_files_to_delete = len(path_sub_directories) - N_USER_KEY_STORED
        for p in path_sub_directories[:n_files_to_delete]:
            user_ids.append(p.name)
            shutil.rmtree(p)

    list_files_tmp = Path("tmp/").iterdir()
    # Delete all files related to user_id
    for file in list_files_tmp:
        for user_id in user_ids:
            if file.name.endswith(f"{user_id}.npy"):
                file.unlink()
mes=[]

def keygen(selected_tasks):
    # Clean tmp directory if needed
    clean_tmp_directory()

    print("Initializing FHEModelClient...")


    
    if not selected_tasks:
        return "choose a task first"  # 修改提示信息为英文
    user_id = numpy.random.randint(0, 2**32)
    if "legal_rating" in selected_tasks:
        model_names.append('legal_rating')
                # Let's create a user_id

        fhe_api= FHEModelClient(FHE_LEGAL_PATH, f".fhe_keys/{user_id}")


    if "financial_rating" in selected_tasks:
        model_names.append('financial_rating')

        fhe_api = FHEModelClient(FHE_MODEL_PATH, f".fhe_keys/{user_id}")
        
        # Let's create a user_id
    
        
    fhe_api.load()
    
        
        # Generate a fresh key
    fhe_api.generate_private_and_evaluation_keys(force=True)
    evaluation_key = fhe_api.get_serialized_evaluation_keys()
    
        # Save evaluation_key in a file, since too large to pass through regular Gradio
        # buttons, https://github.com/gradio-app/gradio/issues/1877
    numpy.save(f"tmp/tmp_evaluation_key_{user_id}.npy", evaluation_key)

    return [list(evaluation_key)[:ENCRYPTED_DATA_BROWSER_LIMIT], user_id]




    
def encode_quantize_encrypt(text, user_id):
    if not user_id:
        raise gr.Error("You need to generate FHE keys first.")
    if "legal_rating" in model_names:
        fhe_api = FHEModelClient(FHE_LEGAL_PATH, f".fhe_keys/{user_id}")
        encodings =vectorizer.fit_transform([text]).toarray()
        if encodings.shape[1] < 1736:
            # 在后面填充零
            padding = np.zeros((1, 1736 - encodings.shape[1]))
            encodings = np.hstack((encodings, padding))
        elif encodings.shape[1] > 1736:
            # 截取前1736列
            encodings = encodings[:, :1736]
    else:
        fhe_api = FHEModelClient(FHE_MODEL_PATH, f".fhe_keys/{user_id}")
        encodings = transformer_vectorizer.transform([text])
    
    fhe_api.load()
    quantized_encodings = fhe_api.model.quantize_input(encodings).astype(numpy.uint8)
    encrypted_quantized_encoding = fhe_api.quantize_encrypt_serialize(encodings)

    # Save encrypted_quantized_encoding in a file, since too large to pass through regular Gradio
    # buttons, https://github.com/gradio-app/gradio/issues/1877
    numpy.save(f"tmp/tmp_encrypted_quantized_encoding_{user_id}.npy", encrypted_quantized_encoding)

    # Compute size
    encrypted_quantized_encoding_shorten = list(encrypted_quantized_encoding)[:ENCRYPTED_DATA_BROWSER_LIMIT]
    encrypted_quantized_encoding_shorten_hex = ''.join(f'{i:02x}' for i in encrypted_quantized_encoding_shorten)
    return (
        encodings[0],
        quantized_encodings[0],
        encrypted_quantized_encoding_shorten_hex,
    )



def run_fhe(user_id):
    encoded_data_path = Path(f"tmp/tmp_encrypted_quantized_encoding_{user_id}.npy")
    if not user_id:
        raise gr.Error("You need to generate FHE keys first.")
    if not encoded_data_path.is_file():
        raise gr.Error("No encrypted data was found. Encrypt the data before trying to predict.")

    # Read encrypted_quantized_encoding from the file
    encrypted_quantized_encoding = numpy.load(encoded_data_path)

    # Read evaluation_key from the file
    evaluation_key = numpy.load(f"tmp/tmp_evaluation_key_{user_id}.npy")

    # Use base64 to encode the encodings and evaluation key
    encrypted_quantized_encoding = base64.b64encode(encrypted_quantized_encoding).decode()
    encoded_evaluation_key = base64.b64encode(evaluation_key).decode()

    query = {}
    query["evaluation_key"] = encoded_evaluation_key
    query["encrypted_encoding"] = encrypted_quantized_encoding
    headers = {"Content-type": "application/json"}
    if "legal_rating" in model_names:
            response = requests.post(
            "http://localhost:8000/predict_legal", data=json.dumps(query), headers=headers
        )
    else:
        response = requests.post(
                "http://localhost:8000/predict_sentiment", data=json.dumps(query), headers=headers
            )
    encrypted_prediction = base64.b64decode(response.json()["encrypted_prediction"])

    # Save encrypted_prediction in a file, since too large to pass through regular Gradio
    # buttons, https://github.com/gradio-app/gradio/issues/1877
    numpy.save(f"tmp/tmp_encrypted_prediction_{user_id}.npy", encrypted_prediction)
    encrypted_prediction_shorten = list(encrypted_prediction)[:ENCRYPTED_DATA_BROWSER_LIMIT]
    encrypted_prediction_shorten_hex = ''.join(f'{i:02x}' for i in encrypted_prediction_shorten)
    return encrypted_prediction_shorten_hex


def decrypt_prediction(user_id):
    encoded_data_path = Path(f"tmp/tmp_encrypted_prediction_{user_id}.npy")
    if not user_id:
        raise gr.Error("You need to generate FHE keys first.")
    if not encoded_data_path.is_file():
        raise gr.Error("No encrypted prediction was found. Run the prediction over the encrypted data first.")

    # Read encrypted_prediction from the file
    encrypted_prediction = numpy.load(encoded_data_path).tobytes()

    if "legal_rating" in model_names:
        fhe_api = FHEModelClient(FHE_LEGAL_PATH, f".fhe_keys/{user_id}")

    fhe_api = FHEModelClient(FHE_MODEL_PATH, f".fhe_keys/{user_id}")
    fhe_api.load()

    # We need to retrieve the private key that matches the client specs (see issue #18)
    fhe_api.generate_private_and_evaluation_keys(force=False)

    predictions = fhe_api.deserialize_decrypt_dequantize(encrypted_prediction)
    print(predictions)
    
    return {
        "low_relative": predictions[0][0],
        "medium_relative": predictions[0][1],
        "high_relative": predictions[0][2],
    }


demo = gr.Blocks()


print("Starting the demo...")
with demo:

    gr.Markdown(
        """

<h2 align="center">📄Cipher Clause</h2>
        <p align="center">
            <img width=600 src="https://www.helloimg.com/i/2024/09/28/66f7f6701bcfb.jpeg">
        </p>

"""
    )


    gr.Markdown(
        """
        <p align="center">
        </p>
        <p align="center">
        </p>
        """
    )

    gr.Markdown("## Notes")
    gr.Markdown(
    """
- The private key is used to encrypt and decrypt the data and shall never be shared.
- The evaluation key is a public key that the server needs to process encrypted data.
"""
    )
    gr.Markdown(
        """
    <hr/>
        """
    )
    gr.Markdown("# Step 0: Select Task")    
    task_checkbox = gr.CheckboxGroup(
                choices=["legal_rating", "financial_rating"],
                label="select_tasks"
            )
    gr.Markdown(
        """
    <hr/>
        """
    )
    gr.Markdown("# Step 1: Generate the keys")

    b_gen_key_and_install = gr.Button("Generate all the keys and send public part to server")

    evaluation_key = gr.Textbox(
        label="Evaluation key (truncated):",
        max_lines=4,
        interactive=False,
    )

    user_id = gr.Textbox(
        label="",
        max_lines=4,
        interactive=False,
        visible=False
    )
    gr.Markdown(
        """
<hr/>
        """
    )
    gr.Markdown("# Step 2: Provide a contract or clause")
    gr.Markdown("## Client side")
    gr.Markdown(
        "Enter a contract or clause you want to analysis)."
    )
    input_type = gr.Radio(choices=["Text Input", "File Upload"], label="Select Input Method")
    user_input = gr.Textbox(label="Enter some words:",visible=False,value="The Employee is entitled to two weeks of paid vacation annually, to be scheduled at the mutual convenience of the Employee and Employer.")

    file_upload = gr.File(label="Upload File", file_types=[".jpg", ".png",".txt"], visible=False)  # Initially hidden
    input_type.change(toggle_visibility, inputs=input_type, outputs=[user_input, file_upload])

    submit_button = gr.Button("Submit")
    text = gr.Textbox(label="Extracted Text")

    submit_button.click(process_input, inputs=[input_type, user_input, file_upload], outputs=text)

    gr.Markdown(
        """
<hr/>
        """
    )
    gr.Markdown("# Step 3: Encode the message with the private key")
    b_encode_quantize_text = gr.Button(
        "Encode, quantize and encrypt the text with vectorizer, and send to server"
    )

    with gr.Row():
        encoding = gr.Textbox(
            label="Representation:",
            max_lines=4,
            interactive=False,
        )
        quantized_encoding = gr.Textbox(
            label="Quantized  representation:", max_lines=4, interactive=False
        )
        encrypted_quantized_encoding = gr.Textbox(
            label="Encrypted quantized representation (truncated):",
            max_lines=4,
            interactive=False,
        )
    gr.Markdown(
        """
<hr/>
        """
    )
    gr.Markdown("# Step 4: Run the FHE evaluation")
    gr.Markdown("## Server side")
    gr.Markdown(
        "The encrypted value is received by the server. Thanks to the evaluation key and to FHE, the server can compute the (encrypted) prediction directly over encrypted values. Once the computation is finished, the server returns the encrypted prediction to the client."
    )

    b_run_fhe = gr.Button("Run FHE execution there")
    encrypted_prediction = gr.Textbox(
        label="Encrypted prediction (truncated):",
        max_lines=4,
        interactive=False,
    )
    gr.Markdown(
        """
<hr/>
        """
    )
    gr.Markdown("# Step 5: Decrypt the class")
    gr.Markdown("## Client side")
    gr.Markdown(
        "The encrypted sentiment is sent back to client, who can finally decrypt it with its private key. Only the client is aware of the original tweet and the prediction."
    )
    b_decrypt_prediction = gr.Button("Decrypt prediction")

    labels_sentiment = gr.Label(label="level:")

    # Button for key generation
    b_gen_key_and_install.click(keygen, inputs=[task_checkbox], outputs=[evaluation_key, user_id])

    # Button to quantize and encrypt
    b_encode_quantize_text.click(
        encode_quantize_encrypt,
        inputs=[text, user_id],
        outputs=[
            encoding,
            quantized_encoding,
            encrypted_quantized_encoding,
        ],
    )

    # Button to send the encodings to the server using post at (localhost:8000/predict_sentiment)
    b_run_fhe.click(run_fhe, inputs=[user_id], outputs=[encrypted_prediction])

    # Button to decrypt the prediction on the client
    b_decrypt_prediction.click(decrypt_prediction, inputs=[user_id], outputs=[labels_sentiment])
    gr.Markdown(
        "The app was built with [Concrete-ML](https://github.com/zama-ai/concrete-ml), a Privacy-Preserving Machine Learning (PPML) open-source set of tools by [Zama](https://zama.ai/). Try it yourself and don't forget to star on Github &#11088;."
    )
demo.launch(share=False)