Privacy considerations in ML Python - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding Differential Privacy

Which of the following best describes the main goal of differential privacy in machine learning?

A To speed up the training process by reducing data size

B To prevent the model from revealing whether any single individual's data was used during training

C To ensure that the model's predictions are always 100% accurate on training data

D To increase the complexity of the model to avoid overfitting

💡 Hint

Think about protecting individual data privacy when training models.

Differential privacy aims to protect individual data by making it hard to tell if any one person's data was included in the training set.

❓ Predict Output

intermediate

Output of Privacy-Preserving Noise Addition

What is the output of the following Python code that adds Laplace noise to a data point for privacy?

ML Python

import numpy as np
np.random.seed(0)
data_point = 10
noise = np.random.laplace(loc=0.0, scale=1.0)
private_data = data_point + noise
print(round(private_data, 2))

A 11.43

B 11.76

C 11.54

D 11.65

💡 Hint

Run the code to see the exact noise value added.

The Laplace noise generated with seed 0 and scale 1.0 added to 10 results in approximately 11.43 after rounding.

❓ Model Choice

advanced

Choosing a Privacy-Preserving Model Technique

You want to train a machine learning model on sensitive health data while ensuring privacy. Which technique is best suited for this?

A Use federated learning to train models locally on devices without sharing raw data

B Train a deep neural network on raw data without modifications

C Collect all data centrally and anonymize by removing names only

D Use a simple linear regression without any privacy measures

💡 Hint

Think about methods that keep data on user devices.

Federated learning trains models locally and only shares updates, protecting raw data privacy.

❓ Metrics

advanced

Evaluating Privacy-Utility Tradeoff

In a privacy-preserving model, which metric combination best reflects a good balance between privacy and model usefulness?

A High epsilon value in differential privacy and low model accuracy

B High epsilon value in differential privacy and high model accuracy

C Low epsilon value in differential privacy and high model accuracy

D Low epsilon value in differential privacy and low model accuracy

💡 Hint

Lower epsilon means stronger privacy.

A low epsilon means stronger privacy, and high accuracy means the model is still useful.

🔧 Debug

expert

Debugging Privacy Leakage in Model Training

You trained a model with differential privacy but found it still leaks some private info. Which code snippet below is the cause?

ML Python

def train_model(data, epsilon):
    # Missing noise addition step
    model = SomeModel()
    model.fit(data)
    return model

A The model uses federated learning, which always leaks data

B The model uses too much noise, causing privacy leakage

C The model trains on encrypted data, which causes leakage

D The model does not add noise to gradients or parameters, so privacy is not enforced

💡 Hint

Differential privacy requires adding noise during training.

Without adding noise, differential privacy is not applied, so private info can leak.