On Fast Adversarial Robustness Adaptation in Model-Agnostic Meta-Learning

Authors: Ren Wang, Kaidi Xu, Sijia Liu, Pin-Yu Chen, Tsui-Wei Weng, Chuang Gan, Meng Wang

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, extensive experiments are conducted to demonstrate the effectiveness of our proposed methods in robust few-shot learning. Codes are available at https://github.com/wangren09/MetaAdv.
Researcher Affiliation Collaboration Ren Wang1,4 Kaidi Xu2 Sijia Liu3,5 Pin-Yu Chen3 Tsui-Wei Weng3 Chuang Gan3 1Rensselaer Polytechnic Institute, USA 2Northeastern University, USA 3MIT-IBM Watson AI Lab, IBM Research, USA 4University of Michigan, USA 5Michigan State University, USA
Pseudocode Yes Algorithm S1 R-MAMLout
Open Source Code Yes Codes are available at https://github.com/wangren09/MetaAdv.
Open Datasets Yes To test the effectiveness of our methods, we employ the Mini Image Net dataset Vinyals et al. (2016), which is the benchmark for few-shot learning.
Dataset Splits Yes Mini Image Net contains 100 classes with 600 samples in each class. We use the training set with 64 classes and test set with 20 classes. ... For the meta-update, we use 15 validation images for each class.
Hardware Specification No The paper does not provide specific details about the hardware used for running experiments, such as GPU models, CPU types, or memory specifications.
Software Dependencies No The paper does not specify the versions of software dependencies, libraries, or frameworks used for the implementation or experiments.
Experiment Setup Yes By default, we set the training attack strength ϵ = 2, γCL = 0.1, and set γout = 5 (TRADES), γout = 0.2 (AT) via a grid search. During meta-testing, a 10-step PGD attack with attack strength ϵ = 2 is used to evaluate RA of the learnt meta-model over 2400 few-shot test tasks. ... We set the gradient step size in the fine-tuning as α = 0.01, and the gradient step sizes in the meta-update as β1 = 0.001, β2 = 0.001 for clean validation data and adversarial validation data, respectively.