Adversarial Robustness in Multi-Task Learning: Promises and Illusions
Authors: Salah Ghamizi, Maxime Cordy, Mike Papadakis, Yves Le Traon697-705
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | we evaluate the design choices that impact the robustness of multi-task deep learning networks. We provide evidence that blindly adding auxiliary tasks, or weighing the tasks provides a false sense of robustness. ... Surprisingly, our experimental study shows that adding more tasks does not consistently increase robustness, and may even have negative effects. |
| Researcher Affiliation | Academia | Salah Ghamizi, Maxime Cordy, Mike Papadakis, and Yves Le Traon University of Luxembourg salah.ghamizi@uni.lu, maxime.cordy@uni.lu, michail.papadakis@uni.lu, yves.letraon@uni.lu |
| Pseudocode | Yes | We describe full algorithm in Appendix D. |
| Open Source Code | Yes | We provide the appendix, all our algorithms, models, and open source-code at https://github.com/yamizi/taskaugment |
| Open Datasets | Yes | We use the Taskonomy dataset, an established dataset for multi-task learning (Zamir et al. 2018). |
| Dataset Splits | Yes | We use the architectures and training settings of the original Taskonomy paper (Zamir et al. 2018) |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU models, or memory specifications used for running the experiments. It only mentions 'Resnet18 encoder' and 'Xception, Wide-Resnet, and Resnet' which are model architectures. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies or libraries used for the experiments. |
| Experiment Setup | Yes | We use as base setting the l Projected Gradient Descent attack (PGD) (Madry et al. 2017) with 25 steps attacks, a strength of ϵ = 8/255 and a step size α = 2/255. ... We use a uniform weights, Cross-entropy loss for the semantic segmentation task and an L1 loss for the other tasks. |