Multitask Learning Strengthens Adversarial Robustness
arXiv:2007.07236
Abstract
Although deep networks achieve strong accuracy on a range of computer vision benchmarks, they remain vulnerable to adversarial attacks, where imperceptible input perturbations fool the network. We present both theoretical and empirical analyses that connect the adversarial robustness of a model to the number of tasks that it is trained on. Experiments on two datasets show that attack difficulty increases as the number of target tasks increase. Moreover, our results suggest that when models are trained on multiple tasks at once, they become more robust to adversarial attacks on individual tasks. While adversarial defense remains an open challenge, our results suggest that deep networks are vulnerable partly because they are trained on too few tasks.
References in corpus (17)
- Intriguing properties of neural networks
- Theoretically Principled Trade-off between Robustness and Accuracy
- GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks
- Defense-GAN: Protecting Classifiers Against Adversarial Attacks Using Generative Models
- One Model To Learn Them All
- Matrix Completion from Noisy Entries
- Houdini: Fooling Deep Structured Prediction Models
- Gradient Surgery for Multi-Task Learning
- Deep Defense: Training DNNs with Improved Adversarial Robustness
- Feature Denoising for Improving Adversarial Robustness
- Evaluating and Understanding the Robustness of Adversarial Logit Pairing
- The Limitations of Deep Learning in Adversarial Settings
- Multi-task Sequence to Sequence Learning
- First-order Adversarial Vulnerability of Neural Networks and Input Dimension
- AdvSPADE: Realistic Unrestricted Attacks for Semantic Segmentation
- Generalization in multitask deep neural classifiers: a statistical physics approach
- Live Trojan Attacks on Deep Neural Networks