Towards Deep Learning Models Resistant to Adversarial Attacks.
0 likes231 views
This document discusses approaches to training deep neural networks to be robust against adversarial examples. It frames adversarial robustness as a minimax game between the network and an attacker. It presents projected gradient descent (PGD) and the Fast Gradient Sign Method (FGSM) as ways to solve the inner maximization problem during training. Experiments show that adversarially trained models can achieve increased robustness compared to standard networks.
1 of 18
Download to read offline
More Related Content
What's hot (20)
Ad
Similar to Towards Deep Learning Models Resistant to Adversarial Attacks. (20)
Ad
More from SEMINARGROOT (20)
Recently uploaded (20)
Towards Deep Learning Models Resistant to Adversarial Attacks.
- 1. Towards Deep Learning Models Resistant to Adversarial Attacks.
- 3. Introduction In particular, resistance to adversarially chosen inputs is becoming a crucial design goal
- 4. Introduction
- 5. Introduction How can we train deep neural networks that are robust to adversarial inputs?
- 6. 2. An Optimization View on Adversarial Robustness An Optimization View on Adversarial Robustness
- 7. An Optimization View on Adversarial Robustness Adversarial attack problem Inner maximization problem Outer minimization problem Min max game Saddle point problem
- 8. An Optimization View on Adversarial Robustness PGD (projected gradient descent ) FGSM (Fast Gradient Sign Method) Inner maximization problem
- 9. 3. Towards Universally Robust Networks Towards Universally Robust Networks
- 10. Towards Universally Robust Networks
- 11. Towards Universally Robust Networks Blue : loss on a standard network Red : adversarially trained model
- 12. 4. Network Capacity and Adversarial Robustness Network Capacity and Adversarial Robustness
- 13. Network Capacity and Adversarial Robustness
- 14. Experiments
- 16. Experiments
- 18. references Towards Deep Learning Models Resistant to Adversarial Attacks, Aleksander Madry, Aleksandar Makelov; ICLR, 2018, Explaining and Harnessing Adversarial Examples, Ian J. Goodfellow, Jonathon Shlens, Christian Szegedy, ICLR, 2015,