Privacy-Preserving AI

Privacy-Preserving AI Techniques

# Differential Privacy (DP)
# Add calibrated noise to protect individual data
from diffprivlib.models import GaussianNaiveBayes
import numpy as np

# Train with differential privacy guarantee
dp_model = GaussianNaiveBayes(epsilon=1.0)  # privacy budget
dp_model.fit(X_train, y_train)

# TensorFlow Privacy
import tensorflow_privacy
from tensorflow_privacy.privacy.optimizers import dp_optimizer_keras

optimizer = dp_optimizer_keras.DPKerasSGDOptimizer(
    l2_norm_clip=1.0,
    noise_multiplier=0.5,
    num_microbatches=32
)

# Federated Learning
# Train on edge devices, only share gradients
# Data never leaves the device
import flwr as fl  # Flower federated learning

class FederatedClient(fl.client.NumPyClient):
    def fit(self, parameters, config):
        model.set_weights(parameters)
        model.fit(local_X, local_y, epochs=1)
        return model.get_weights(), len(local_X), {}

# Secure Multi-Party Computation (MPC)
# Multiple parties compute on combined data without seeing each other's data

# Homomorphic Encryption
# Compute on encrypted data without decrypting
import tenseal as ts