watermarking.py

from sklearn.model_selection import train_test_split
import pandas as pd
from sklearn import datasets
from sklearn.preprocessing import MinMaxScaler
import pandas as pd
import numpy as np
from concrete.ml.sklearn import SGDClassifier
import time
from concrete.ml.sklearn import NeuralNetClassifier
import hashlib

RANDOM_STATE = 6

np.random.seed(RANDOM_STATE) #2 #1


def gen_database():
    rng = np.random.default_rng(42)

    X, y = datasets.load_breast_cancer(return_X_y=True)
    x_train, x_test, y_train, y_test = train_test_split(X, y, test_size=0.3, stratify=y)

    scaler = MinMaxScaler(feature_range=[-1, 1])
    x_train = scaler.fit_transform(x_train)
    x_test = scaler.transform(x_test)

    perm = rng.permutation(x_train.shape[0])

    x_train = x_train[perm, ::]
    y_train = y_train[perm]

    return x_train, y_train, x_test, y_test

def gen_trigger_set():
    X_trigger = np.random.random_sample((15, 30))
    y_trigger = np.random.randint(0, 2, (15))
    for i in range(15):
        if y_trigger[i] == 1:
            X_trigger[i, :15] = X_trigger[i, 15]
        else:
            X_trigger[i, 15:] = X_trigger[i, 15]
    return X_trigger, y_trigger


def train_model(x_train, y_train):

    parameters_range = (-1.0, 1.0)
    model = SGDClassifier(
        random_state=42,
        max_iter=100,
        fit_encrypted=True,
        parameters_range=parameters_range,
        penalty=None,
        learning_rate="constant",
        verbose=1)
    # %%
    model.fit(x_train, y_train, fhe="simulate")
    return model

def watermark_model(model, X_trigger, y_trigger):
    model.max_iter = 17
    model.alpha = 1e-6
    model.penalty = "l2"
    model.warm_start = True

    a = time.time()
    model.fit(X_trigger, y_trigger, fhe="simulate")
    print("Time :", time.time() - a)

    return model

def evaluate(model, x_train, y_train, x_test, y_test, X_trigger, y_trigger):
    print(f"Accuracy Train Set :{np.sum(model.predict(x_train) == y_train) / len(y_train)}")
    print(f"Accuracy Test Set :{np.sum(model.predict(x_test) == y_test) / len(y_test)}")
    print(f"Accuracy Trigger Set :{np.sum(model.predict(X_trigger) == y_trigger) / len(y_trigger)}")


def get_model_hash(model):
    m = hashlib.sha256()
    m.update(model.coef_)
    m.hexdigest()
    return m.hexdigest()

def get_trigger_hash(X_trigger, y_trigger):
    y_trigger = y_trigger.reshape(-1, 1)
    trigger_set = np.concatenate((X_trigger, y_trigger), axis=1)

    m = hashlib.sha256()
    m.update(trigger_set)
    m.hexdigest()

    return m.hexdigest()

def test():

    # Gen data
    x_train, y_train, x_test, y_test =  gen_database()

    np.save("x_train", x_train)
    np.save("y_train", y_train)
    np.save("x_test", x_test)
    np.save("y_test", y_test)

    X_trigger, y_trigger = gen_trigger_set()

    np.save("x_trigger", X_trigger)
    np.save("y_trigger", y_trigger)

    X_trigger, y_trigger = np.load("x_trigger.npy"), np.load("y_trigger.npy")

    model =  train_model(x_train, y_train)

    np.save("model_coef", model.coef_)
    np.save("model_intercept", model.intercept_)

    model.coef_ = np.load("model_coef.npy")
    model.intercept_ = np.load("model_intercept.npy")

    wat_model = watermark_model(model, X_trigger, y_trigger)

    np.save("wat_model_coef", wat_model.coef_)
    np.save("wat_model_intercept", wat_model.intercept_)

    wat_model.coef_ = np.load("wat_model_coef.npy")
    wat_model.intercept_ = np.load("wat_model_intercept.npy")

    evaluate(wat_model, x_train, y_train, x_test, y_test, X_trigger, y_trigger)

# test()