Security Vulnerabilities in AI Assistant Based Applications

Overview

Explore security vulnerabilities in AI assistant-based applications through this AppSecUSA 2018 conference talk by Abraham Kang. Delve into the world of intelligent assistants, learning how they can be compromised despite seemingly secure setups. Discover various attack vectors, including physical real-world attacks, splicing techniques, and future potential threats. Gain insights into the architecture of AI assistants, understanding slots and their vulnerabilities. Examine neural networks and techniques for attacking them, including adversarial examples, masks, and patches. Learn about white box and black box adversarial attacks, as well as methods for defending against these threats. Investigate trojaning neural networks, model and training data extraction, and receive a comprehensive summary of AI assistant security concerns. Equip yourself with the knowledge to identify and address vulnerabilities in AI assistant applications.

Syllabus

Intro
Typical Setup at Home
Physical Real-World Attacks
Splicing Demo 1
Splicing Demo 2
Future Attacks 1
Future Attacks 2
Attacking Al Assistant Business Logic
Architecture
Understanding Slots
Attackable Slots
Neural Networks and the Brain
Techniques for Attacking Neural Networks
What Can You Attack with Adversarial Examples?
Why Do Adversarial Masks work?
Adversarial Result
Adversarial Input Generation Techniques
White Box Adversarial Attack • Techniques
White Box Adversarial Attack Techniques
Black Box Adversarial Attack
Adversarial Patches
Defending against adversarial samples
Trojaning neural networks
Defending against trojans
Model Data Extraction
Training Data Extraction
Summary