Strengthening Deep Neural Networks by Warr, Katy -- Read -- Imperial Library of Trantor

Index

Preface

Who Should Read This Book How This Book Is Organized Conventions Used in This Book Using Code Examples The Mathematics in This Book O’Reilly Online Learning How to Contact Us Acknowledgments

I. An Introduction to Fooling AI 1. Introduction

A Shallow Introduction to Deep Learning A Very Brief History of Deep Learning AI “Optical Illusions”: A Surprising Revelation What Is “Adversarial Input”?

Adversarial Perturbation Unnatural Adversarial Input Adversarial Patches Adversarial Examples in the Physical World

The Broader Field of “Adversarial Machine Learning” Implications of Adversarial Input

2. Attack Motivations

Circumventing Web Filters Online Reputation and Brand Management Camouflage from Surveillance Personal Privacy Online Autonomous Vehicle Confusion Voice Controlled Devices

3. Deep Neural Network (DNN) Fundamentals

Machine Learning A Conceptual Introduction to Deep Learning DNN Models as Mathematical Functions

DNN Inputs and Outputs DNN Internals and Feed-Forward Processing How a DNN Learns

Creating a Simple Image Classifier

4. DNN Processing for Image, Audio, and Video

Image

Digital Representation of Images DNNs for Image Processing Introducing CNNs

Audio

Digital Representation of Audio DNNs for Audio Processing Introducing RNNs Speech Processing

Video

Digital Representation of Video DNNs for Video Processing

Adversarial Considerations Image Classification Using ResNet50

II. Generating Adversarial Input 5. The Principles of Adversarial Input

The Input Space

Generalizations from Training Data Experimenting with Out-of-Distribution Data

What’s the DNN Thinking? Perturbation Attack: Minimum Change, Maximum Impact Adversarial Patch: Maximum Distraction Measuring Detectability

A Mathematical Approach to Measuring Perturbation Considering Human Perception

Summary

6. Methods for Generating Adversarial Perturbation

White Box Methods

Searching the Input Space Exploiting Model Linearity Adversarial Saliency Increasing Adversarial Confidence Variations on White Box Approaches

Limited Black Box Methods Score-Based Black Box Methods Summary

III. Understanding the Real-World Threat 7. Attack Patterns for Real-World Systems

Attack Patterns

Direct Attack Replica Attack Transfer Attack Universal Transfer Attack

Reusable Patches and Reusable Perturbation Bringing It Together: Hybrid Approaches and Trade-offs

8. Physical-World Attacks

Adversarial Objects

Object Fabrication and Camera Capabilities Viewing Angles and Environment

Adversarial Sound

Audio Reproduction and Microphone Capabilities Audio Positioning and Environment

The Feasibility of Physical-World Adversarial Examples

IV. Defense 9. Evaluating Model Robustness to Adversarial Inputs

Adversarial Goals, Capabilities, Constraints, and Knowledge

Goals Capabilities, Knowledge, and Access

Model Evaluation

Empirically Derived Robustness Metrics Theoretically Derived Robustness Metrics

Summary

10. Defending Against Adversarial Inputs

Improving the Model

Gradient Masking Adversarial Training Out-of-Distribution Confidence Training Randomized Dropout Uncertainty Measurements

Data Preprocessing

Preprocessing in the Broader Processing Chain Intelligently Removing Adversarial Content

Concealing the Target Building Strong Defenses Against Adversarial Input

Open Projects Taking a Holistic View

11. Future Trends: Toward Robust AI

Increasing Robustness Through Outline Recognition Multisensory Input Object Composition and Hierarchy Finally…

A. Mathematics Terminology Reference Index

← Prev
Back
Next →

← Prev
Back
Next →