Adversarial Training on Purification (AToP): Advancing Both Robustness and Generalization

Authors: Guang Lin, Chao Li, Jianhai Zhang, Toshihisa Tanaka, Qibin Zhao

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To evaluate our method in an efficient and scalable way, we conduct extensive experiments on CIFAR-10, CIFAR-100, and Image Nette to demonstrate that our method achieves optimal robustness and exhibits generalization ability against unseen attacks.
Researcher Affiliation Academia 1Tokyo University of Agriculture and Technology 2RIKEN Center for Advanced Intelligence Project (RIKEN AIP) 3Hangzhou Dianzi University
Pseudocode Yes Algorithm 1 Adversarial Training on Purification method (ATo P) Require: Training examples x, ground truth y, parameters of classifier model θf, parameters of purifier model θg, training epoch Nep 1: Initialize θf and θg with pre-trained classifier model and pre-trained purifier model. 2: for epoch = 1...Nep do 3: Build adversarial examples x with perturbations δ: x x + δ 4: Freeze θf and update θg with gradient descent based on loss in Eq. (9). 5: θg θg θg 6: end for 7: return purifier model with θg
Open Source Code No The paper does not provide a direct link or explicit statement about the public availability of its source code.
Open Datasets Yes We conduct extensive experiments on CIFAR-10, CIFAR-100 (Krizhevsky et al., 2009) and Image Nette (a subset of 10 classified classes from Image Net) (Howard, 2021) to empirically validate the effectiveness of the proposed methods against adversarial attacks.
Dataset Splits No The paper mentions using CIFAR-10, CIFAR-100, and Image Nette and that 'we randomly select 512 images from the test set for robust evaluation', but does not explicitly state the training, validation, and test dataset splits or percentages used for training and validation, nor does it confirm the use of standard predefined splits for all sets.
Hardware Specification Yes All experiments presented in the paper are conducted under these hyperparameters and performed by NVIDIA RTX A5000 Graphics Card with 24GB GDDR6 GPU memory, CUDA v11.7 and cu DNN v8.5.0 in Py Torch v1.13.11 (Paszke et al., 2019).
Software Dependencies Yes All experiments presented in the paper are conducted under these hyperparameters and performed by NVIDIA RTX A5000 Graphics Card with 24GB GDDR6 GPU memory, CUDA v11.7 and cu DNN v8.5.0 in Py Torch v1.13.11 (Paszke et al., 2019).
Experiment Setup Yes Training details: After experimental testing, we have determined the hyperparameters: Gaussian standard deviation σ = 0.25, mask size p = P/8 (P is side length of input x), mask number N = 4, missing rate r = 0.25 and weight λ = 0.1.