The Bad Guys in AI: Attacking Machine Learning Systems
summary
Introduces generative adversarial networks (GANs), explains how GANs can learn to imitate and fool deep neural networks, and shows why that creates vulnerabilities in machine-learning systems. Includes a Python example demonstrating an adversarial attack against an image-recognition system and closes with a discussion of the security implications for critical ML deployments.
AI-written summary · about
chapters (21)
- 0:00Introduction to attacks
- 1:25Deep neural networks
- 3:11Key gradient descent
- 4:35Generative vs discriminative
- 6:36What is an adversarial attack
- 7:37Targeted vs untargeted attacks
- 8:17Black-box attack steps
- 9:40Demo: food model attack
- 13:10Attack tools
- 14:12Oracle data extraction
- 15:15Training substitute model
- 16:40Synthetic data generation
- 17:01Adversarial attack execution
- 17:43Attack results
- 18:26Real-world implications
- 19:30Current defenses
- 20:11Adversarial training
- 20:32Proactive defenses
- 21:13AI responsibility
- 21:54Societal impact
- 22:35Q&A
related talks
27mAdversarial Intelligence: Using GANs in Cybersecurity
40mOverview of Generative Adversarial Networks (GANs)
47mUnmasking Generative Adversarial Networks
13mNavigating Generative AI and Its Risks
46mA Short Introduction to Hacking Artificial Intelligence
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
45mOverview of Risks, Vulnerabilities and Paths to Safer Use of Generative AI
46mGenerative AI: Cyber resiliency and risk