backend.py

"""Backend functions used in the app."""

import os
import shutil
import gradio as gr
import numpy
import requests
import pickle
import pandas
from itertools import chain

from settings import (
    SERVER_URL,
    FHE_KEYS,
    CLIENT_FILES,
    SERVER_FILES,
    DEPLOYMENT_PATH,
    PROCESSED_INPUT_SHAPE,
    INPUT_INDEXES,
    INPUT_SLICES,
    PRE_PROCESSOR_USER_PATH, 
    PRE_PROCESSOR_BANK_PATH,
    PRE_PROCESSOR_THIRD_PARTY_PATH,
    CLIENT_TYPES,
    USER_COLUMNS,
    BANK_COLUMNS,
    THIRD_PARTY_COLUMNS,
    YEARS_EMPLOYED_BINS,
    YEARS_EMPLOYED_BIN_NAME_TO_INDEX,
)

from utils.client_server_interface import MultiInputsFHEModelClient

# Load pre-processor instances
with (
    PRE_PROCESSOR_USER_PATH.open('rb') as file_user, 
    PRE_PROCESSOR_BANK_PATH.open('rb') as file_bank,
    PRE_PROCESSOR_THIRD_PARTY_PATH.open('rb') as file_third_party,
):
    PRE_PROCESSOR_USER = pickle.load(file_user)
    PRE_PROCESSOR_BANK = pickle.load(file_bank)
    PRE_PROCESSOR_THIRD_PARTY = pickle.load(file_third_party)


def shorten_bytes_object(bytes_object, limit=500):
    """Shorten the input bytes object to a given length.

    Encrypted data is too large for displaying it in the browser using Gradio. This function
    provides a shorten representation of it.

    Args:
        bytes_object (bytes): The input to shorten
        limit (int): The length to consider. Default to 500.

    Returns:
        str: Hexadecimal string shorten representation of the input byte object. 

    """
    # Define a shift for better display
    shift = 100
    return bytes_object[shift : limit + shift].hex()


def clean_temporary_files(n_keys=20):
    """Clean older keys and encrypted files.

    A maximum of n_keys keys and associated temporary files are allowed to be stored. Once this 
    limit is reached, the oldest files are deleted.

    Args:
        n_keys (int): The maximum number of keys and associated files to be stored. Default to 20.

    """
    # Get the oldest key files in the key directory
    key_dirs = sorted(FHE_KEYS.iterdir(), key=os.path.getmtime)

    # If more than n_keys keys are found, remove the oldest
    client_ids = []
    if len(key_dirs) > n_keys:
        n_keys_to_delete = len(key_dirs) - n_keys
        for key_dir in key_dirs[:n_keys_to_delete]:
            client_ids.append(key_dir.name)
            shutil.rmtree(key_dir)
    
    # Delete all files related to the IDs whose keys were deleted
    for directory in chain(CLIENT_FILES.iterdir(), SERVER_FILES.iterdir()):
        for client_id in client_ids:
            if client_id in directory.name:
                shutil.rmtree(directory)


def _get_client(client_id):
    """Get the client instance.

    Args:
        client_id (int): The client ID to consider.

    Returns:
        FHEModelClient: The client instance.
    """
    key_dir = FHE_KEYS / f"{client_id}"

    return MultiInputsFHEModelClient(DEPLOYMENT_PATH, key_dir=key_dir, nb_inputs=len(CLIENT_TYPES))


def _get_client_file_path(name, client_id, client_type=None):
    """Get the file path for the client.

    Args:
        name (str): The desired file name (either 'evaluation_key', 'encrypted_inputs' or 
            'encrypted_outputs').
        client_id (int): The client ID to consider.
        client_type (Optional[str]): The type of user to consider (either 'user', 'bank', 
            'third_party' or None). Default to None, which is used for evaluation key and output.

    Returns:
        pathlib.Path: The file path.
    """
    client_type_suffix = "" 
    if client_type is not None:
        client_type_suffix = f"_{client_type}"

    dir_path = CLIENT_FILES / f"{client_id}"
    dir_path.mkdir(exist_ok=True)

    return dir_path / f"{name}{client_type_suffix}"


def _send_to_server(client_id, client_type, file_name):
    """Send the encrypted inputs or the evaluation key to the server.

    Args:
        client_id (int): The client ID to consider.
        client_type (Optional[str]): The type of client to consider (either 'user', 'bank', 
            'third_party' or None).
        file_name (str): File name to send (either 'evaluation_key' or 'encrypted_inputs').
    """
    # Get the paths to the encrypted inputs
    encrypted_file_path = _get_client_file_path(file_name, client_id, client_type)

    # Define the data and files to post
    data = {
        "client_id": client_id,
        "client_type": client_type,
        "file_name": file_name,
    }

    files = [
        ("files", open(encrypted_file_path, "rb")),
    ]

    # Send the encrypted inputs or evaluation key to the server
    url = SERVER_URL + "send_file"
    with requests.post(
        url=url,
        data=data,
        files=files,
    ) as response:
        return response.ok


def keygen_send():
    """Generate the private and evaluation key, and send the evaluation key to the server.
    
    Returns:
        client_id (str): The current client ID to consider.
    """
    # Clean temporary files
    clean_temporary_files()

    # Create an ID for the current client to consider
    client_id = numpy.random.randint(0, 2**32)

    # Retrieve the client instance
    client = _get_client(client_id)

    # Generate the private and evaluation keys
    client.generate_private_and_evaluation_keys(force=True)

    # Retrieve the serialized evaluation key
    evaluation_key = client.get_serialized_evaluation_keys()

    file_name = "evaluation_key"

    # Save evaluation key as bytes in a file as it is too large to pass through regular Gradio
    # buttons (see https://github.com/gradio-app/gradio/issues/1877)
    evaluation_key_path = _get_client_file_path(file_name, client_id)

    with evaluation_key_path.open("wb") as evaluation_key_file:
        evaluation_key_file.write(evaluation_key)

    # Send the evaluation key to the server
    _send_to_server(client_id, None, file_name)

    # Create a truncated version of the evaluation key for display
    evaluation_key_short = shorten_bytes_object(evaluation_key)
    
    return client_id, evaluation_key_short, gr.update(value="Keys are generated and evaluation key is sent ✅")


def _encrypt_send(client_id, inputs, client_type):
    """Encrypt the given inputs for a specific client and send it to the server.

    Args:
        client_id (str): The current client ID to consider.
        inputs (numpy.ndarray): The inputs to encrypt.
        client_type (str): The type of client to consider (either 'user', 'bank' or 'third_party').
    
    Returns:
        encrypted_inputs_short (str): A short representation of the encrypted input to send in hex. 
    """
    if client_id == "":
        raise gr.Error("Please generate the keys first.")

    # Retrieve the client instance
    client = _get_client(client_id)

    # Quantize, encrypt and serialize the inputs
    encrypted_inputs = client.quantize_encrypt_serialize_multi_inputs(
        inputs, 
        input_index=INPUT_INDEXES[client_type], 
        processed_input_shape=PROCESSED_INPUT_SHAPE, 
        input_slice=INPUT_SLICES[client_type],
    )

    file_name = "encrypted_inputs"

    # Save encrypted_inputs to bytes in a file, since too large to pass through regular Gradio
    # buttons, https://github.com/gradio-app/gradio/issues/1877
    encrypted_inputs_path = _get_client_file_path(file_name, client_id, client_type)

    with encrypted_inputs_path.open("wb") as encrypted_inputs_file:
        encrypted_inputs_file.write(encrypted_inputs)

    # Create a truncated version of the encrypted inputs for display
    encrypted_inputs_short = shorten_bytes_object(encrypted_inputs)

    _send_to_server(client_id, client_type, file_name)

    return encrypted_inputs_short


def pre_process_encrypt_send_user(client_id, *inputs):
    """Pre-process, encrypt and send the user inputs for a specific client to the server.

    Args:
        client_id (str): The current client ID to consider.
        *inputs (Tuple[numpy.ndarray]): The inputs to pre-process.
    
    Returns:
        (str): A short representation of the encrypted input to send in hex. 
    """
    bool_inputs, num_children, household_size, total_income, age, income_type, education_type, \
        family_status, occupation_type, housing_type = inputs

    # Retrieve boolean values
    own_car = "Car" in bool_inputs
    own_property = "Property" in bool_inputs
    mobile_phone = "Mobile phone" in bool_inputs

    user_inputs = pandas.DataFrame({
        "Own_car": [own_car],
        "Own_property": [own_property],
        "Mobile_phone": [mobile_phone],
        "Num_children": [num_children],
        "Household_size": [household_size],
        "Total_income": [total_income],
        "Age": [age],
        "Income_type": [income_type],
        "Education_type": [education_type],
        "Family_status": [family_status],
        "Occupation_type": [occupation_type],
        "Housing_type": [housing_type],
    })

    user_inputs = user_inputs.reindex(USER_COLUMNS, axis=1)

    preprocessed_user_inputs = PRE_PROCESSOR_USER.transform(user_inputs)

    return _encrypt_send(client_id, preprocessed_user_inputs, "user")


def pre_process_encrypt_send_bank(client_id, *inputs):
    """Pre-process, encrypt and send the bank inputs for a specific client to the server.

    Args:        
        client_id (str): The current client ID to consider.
        *inputs (Tuple[numpy.ndarray]): The inputs to pre-process.
    
    Returns:
        (str): A short representation of the encrypted input to send in hex. 
    """
    account_age = inputs[0]

    bank_inputs = pandas.DataFrame({
        "Account_age": [account_age],
    })

    bank_inputs = bank_inputs.reindex(BANK_COLUMNS, axis=1)

    preprocessed_bank_inputs = PRE_PROCESSOR_BANK.transform(bank_inputs)
    
    return _encrypt_send(client_id, preprocessed_bank_inputs, "bank")


def pre_process_encrypt_send_third_party(client_id, *inputs):
    """Pre-process, encrypt and send the third party inputs for a specific client to the server.

    Args:
        client_id (str): The current client ID to consider.
        *inputs (Tuple[numpy.ndarray]): The inputs to pre-process.

    Returns:
        (str): A short representation of the encrypted input to send in hex. 
    """
    years_employed_bin, employed = inputs

    years_employed = YEARS_EMPLOYED_BIN_NAME_TO_INDEX[years_employed_bin]
    is_employed = employed == "Yes"

    third_party_inputs = pandas.DataFrame({
        "Years_employed": [years_employed],
        "Employed": [is_employed],
    })

    third_party_inputs = third_party_inputs.reindex(THIRD_PARTY_COLUMNS, axis=1)
    preprocessed_third_party_inputs = PRE_PROCESSOR_THIRD_PARTY.transform(third_party_inputs)

    return _encrypt_send(client_id, preprocessed_third_party_inputs, "third_party")


def run_fhe(client_id):
    """Run the model on the encrypted inputs previously sent using FHE.

    Args:
        client_id (str): The current client ID to consider.
    """

    if client_id == "":
        raise gr.Error("Please generate the keys first.")

    data = {
        "client_id": client_id,
    }

    # Trigger the FHE execution on the encrypted inputs previously sent
    url = SERVER_URL + "run_fhe"
    with requests.post(
        url=url,
        data=data,
    ) as response:
        if response.ok:
            return response.json()
        else:
            raise gr.Error("Please send the inputs from all three parties to the server first.")


def get_output_and_decrypt(client_id):
    """Retrieve the encrypted output.

    Args:
        client_id (str): The current client ID to consider.
    
    Returns:
        (Tuple[str, bytes]): The output message based on the decrypted prediction as well as 
            a byte short representation of the encrypted output. 
    """

    if client_id == "":
        raise gr.Error("Please generate the keys first.")

    data = {
        "client_id": client_id,
    }

    # Retrieve the encrypted output
    url = SERVER_URL + "get_output"
    with requests.post(
        url=url,
        data=data,
    ) as response:
        if response.ok:
            encrypted_output_proba = response.content
            
            # Create a truncated version of the encrypted inputs for display
            encrypted_output_short = shorten_bytes_object(encrypted_output_proba)

            # Retrieve the client API
            client = _get_client(client_id)

            # Deserialize, decrypt and post-process the encrypted output
            output_proba = client.deserialize_decrypt_dequantize(encrypted_output_proba)

            # Determine the predicted class
            output = numpy.argmax(output_proba, axis=1).squeeze()
            
            return (
                "Credit card is likely to be approved ✅" if output == 1 
                else "Credit card is likely to be denied ❌",
                encrypted_output_short,
            )

        else:
            raise gr.Error("Please run the FHE execution first and wait for it to be completed.")


def explain_encrypt_run_decrypt(client_id, prediction_output, *inputs):
    """Pre-process and encrypt the inputs, run the prediction in FHE and decrypt the output. 

    Args:
        client_id (str): The current client ID to consider.
        prediction_output (str): The initial prediction output. This parameter is only used to 
            throw an error in case the prediction was positive. 
        *inputs (Tuple[numpy.ndarray]): The inputs to consider.
    
    Returns:
        (str): A message indicating the number of additional years of employment that could be 
            required in order to increase the chance of credit card approval.
    """

    if "approved" in prediction_output:
        raise gr.Error(
            "Explaining the prediction can only be done if the credit card is likely to be denied."
        )

    # Retrieve the third party inputs
    years_employed, employed = inputs

    # Years_employed is divided into several ordered bins. Here, we retrieve the index representing 
    # the bin from the input
    bin_index = YEARS_EMPLOYED_BIN_NAME_TO_INDEX[years_employed]

    # If the bin is not the last (representing the most years of employment), we run the model in 
    # FHE for each bins "older" than the given bin, in order. Then, we retrieve the first bin that
    # changes the model's prediction to "approval" and display it to the user.  
    if bin_index != len(YEARS_EMPLOYED_BINS) - 1:

        output_predictions = []
        
        # Loop over the bins "older" than the input one
        for years_employed_bin in YEARS_EMPLOYED_BINS[bin_index+1:]:
            
            # Send the new encrypted input
            pre_process_encrypt_send_third_party(client_id, years_employed_bin, employed)

            # Run the model in FHE
            run_fhe(client_id)

            # Retrieve the new prediction
            output_prediction = get_output_and_decrypt(client_id)

            is_approved = "approved" in output_prediction[0]
            output_predictions.append(is_approved)
        
        # Re-send the initial third party inputs in order to avoid unwanted conflict (as sending
        # some inputs basically re-writes the associated file on the server side)
        pre_process_encrypt_send_third_party(client_id, years_employed, employed)
        
        # In case the model predicted at least one approval
        if any(output_predictions):

            # Retrieve the first bin that made the model predict an approval 
            first_approved_prediction_index = numpy.argmax(output_predictions)
            years_employed_bin_needed = YEARS_EMPLOYED_BINS[first_approved_prediction_index + bin_index + 1]

            return (
                f"Having at least {years_employed_bin_needed} more years of employment would "
                "increase your chance of having your credit card approved."
            )

        return (
            f"Increasing the number of years of employment up to {YEARS_EMPLOYED_BINS[-1]} years "
            "does not seem to be enough to get an approval based on the given inputs. Other inputs "
            "like the income or the account's age might have bigger impact in this particular case."
        )

    return (
        f"You already have the maximum amount of years of employment ({years_employed} years). "
        "Other inputs like the income or the account's age might have bigger impact in this "
        "particular case."
    )