One Pixel Adversarial Attacks via Sketched Programs
summary
Claims a method called OPPSLA that synthesizes a sketch-based prioritization program to find one-pixel (or few-pixel) adversarial examples for CIFAR-10 and ImageNet using far fewer queries than prior work. The tool uses a Metropolis-Hastings–inspired stochastic search over typed ASTs with conditions based on RGB stats, distance-from-center, and confidence change; synthesis on ~50 training images typically costs ~50,000 queries once, after which attacks achieve state-of-the-art success rates and often an order-of-magnitude fewer queries (59% higher success under a one-query limit versus Spar).
AI-written summary · about
chapters (9)
- 0:00One pixel adversarial attacks
- 1:06Challenges in one pixel attacks
- 3:16Problem definition and goals
- 4:01Sketch programs for attacks
- 7:15Dynamic prioritization conditions
- 9:05Synthesizer and stochastic search
- 11:14Evaluation vs state-of-the-art
- 12:41Summary of contributions
- 13:23Q&A
related talks
21mAdversarial Attacks
38mUsing adversarial samples to robustify your Neural Network Models
25mTricking Neural Networks : Explore Adversarial Attacks
36mAdversarial Machine Learning: How Secure Is ML?
55mPerception Deception: Physical Adversarial Attack Challenges
20mAttacking and Defending Machine Learning Applications of Public Cloud
36mAdversarial Attacks on Deep Learning
34mHacking the Blackbox