2019 Triangle Machine Learning Day - Mitigating Evasion Attacks to Deep Neural Networks via Region-based Classification - Xiaoyu Cao, September 20, 2019
Download as PPTX, PDF1 like151 views
The document discusses the challenge of mitigating evasion attacks on deep neural networks, where perturbations are added to inputs to deceive classifiers. It reviews existing defense methods that either compromise classification accuracy or require manual processing of detected adversarial examples. The authors propose a region-based classification approach to improve robustness against targeted attacks while evaluating its effectiveness on datasets like MNIST and CIFAR-10.
Related topics:1 of 6
Download to read offline
More Related Content
Similar to 2019 Triangle Machine Learning Day - Mitigating Evasion Attacks to Deep Neural Networks via Region-based Classification - Xiaoyu Cao, September 20, 2019 (20)
Ad
More from The Statistical and Applied Mathematical Sciences Institute (20)
Ad
2019 Triangle Machine Learning Day - Mitigating Evasion Attacks to Deep Neural Networks via Region-based Classification - Xiaoyu Cao, September 20, 2019
- 1. Mitigating Evasion Attacks to Deep Neural Networks via Region-based Classification Xiaoyu Cao, Neil Zhenqiang Gong 1
- 2. Security is a Big Challenge for Deploying Deep Neural Networks 2 • Evasion Attacks • Perturbation is added to the input at test time to fool the classifiers. * Image from Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. ICLR 2015.
- 3. Existing Defenses 3 • New methods to train neural networks • Sacrifice classification accuracy on benign examples. • Detecting adversarial examples • Need human to manually process the detected adversarial examples. No benefit of automatic decision makings.
- 4. Understanding Adversarial Examples 4 𝑅𝑖 𝑅𝑡 Classification boundary of true class 𝒊 and target class 𝒕
- 5. Design Region-based Classification 5 Input Output 5 … … 0 5 … 0 0 Ensemble
- 6. Experimental Results 6 TABLE : Classification accuracy on benign examples and robustness to targeted CW attacks on (𝑎)MNIST dataset (𝑏) CIFAR-10 dataset. (𝑎) (𝑏)