Differential Privacy in Responsible AI

Privacy Enhancing Techniques: Overview

Synthetic data

Federated learning

Pseudonymization

Artificially generated data for a given use case instead of the data captured directly.

The data owner allows the system to use it for insights without sharing the actual data.

Artificial identifiers replace PTI fields within the dataset.

Generative Adversarial Networks (GANs)

Homomorphic encryption

Differential privacy

Sensitive data is converted to Ciphertext (plain text transformed using an encryption algorithm).

Competing neural networks attempt to become more accurate than others.

A degree of randomization is added to the dataset to maintain individuals’ privacy. Since the amount of noise added gets controlled, generated aggregate insights are still accurate.

Options To Apply Differential Privacy to a Machine Learning Workflow

Adding noise during data collection

Adding noise to the data set

Train a non-private baseline model for comparison

Adding noise during aggregation

A key question in selecting the best approach is which stakeholders should be allowed to access the data in an unprotected state.

2

© 2023 Fractal Analytics Inc. All rights reserved

Made with FlippingBook - PDF hosting