Adversarial Attacks on Deep Learning
summary
Adversarial perturbations reliably flip classifiers with tiny, often imperceptible changes and are pervasive across architectures and non‑deep models; optimization-based attacks and cheap fast gradient sign method attacks both work, white‑box yields highest success, black‑box still surprisingly effective, and some attacks are transferable or universal. Real‑world proofs — printed stickers, graffiti on signs, 3D printed adversarial object — successfully fool cameras. Attacks on autoencoders are harder; defenses (including adversarial training) help but no reliable, general solution exists.
AI-written summary · about
chapters (8)
- 0:00Introduction to adversarial attacks
- 1:02What are adversarial attacks
- 4:21Real-world adversarial attacks
- 8:14Taxonomy of attack types
- 11:29Technical deep dive into attacks
- 20:52Attacking autoencoders
- 25:12Defenses and current state
- 27:52Q&A session
related talks
21mAdversarial Attacks
36mAdversarial Machine Learning: How Secure Is ML?
44mAdversarial attacks on deep learning models
38mUsing adversarial samples to robustify your Neural Network Models
02h 13mUnderstanding Security Threats Against Machine/Deep Learning Applications
46mA Short Introduction to Hacking Artificial Intelligence
25mTricking Neural Networks : Explore Adversarial Attacks
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence