ImageNet Pre-training Also Transfers Non-robustness
Authors: Jiaming Zhang, Jitao Sang, Qi Yi, Yunfan Yang, Huiwen Dong, Jian Yu
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We first conducted experiments on various datasets and network backbones to uncover the adversarial non-robustness in fine-tuned model. Further analysis was conducted on examining the learned knowledge of fine-tuned model and standard model, and revealed that the reason leading to the nonrobustness is the non-robust features transferred from Image Net pre-trained model. Table 1: Comparison of generalization and robustness between standard model, partial fine-tuned model and full fine-tuned model. For each model, we report accuracy of original inputs, accuracy of adversarial inputs, and decline ratio. |
| Researcher Affiliation | Academia | Jiaming Zhang1, Jitao Sang1,2*, Qi Yi1, Yunfan Yang1, Huiwen Dong3, Jian Yu1 1School of Computer and Information Technology, Beijing Jiaotong University 2Peng Cheng Lab 3BNU-UIC Institute of Artificial Intelligence and Future Networks, Beijing Normal University {jiamingzhang, jtsang, 21125273, 19281298}@bjtu.edu.cn, 202131081024@mail.bnu.edu.cn, jianyu@bjtu.edu.cn |
| Pseudocode | No | The paper describes methods and equations, such as Equation (5) for the objective function, but does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/jiamingzhang94/Image Net-Pretraining-transfers-non-robustness. |
| Open Datasets | Yes | We carry out our study on several widely-used image classification datasets including Pets (Parkhi et al. 2012), NICO (He, Shen, and Cui 2020), Flowers (Nilsback and Zisserman 2008), Cars (Krause et al. 2013), Food (Bossard, Guillaumin, and Van Gool 2014), and CIFAR10 (Krizhevsky, Hinton et al. 2009). |
| Dataset Splits | No | The paper mentions '1, 000 training images and 200 testing images for each letter class' for the Alphabet dataset, but does not explicitly provide details for a validation split for any of the datasets used, nor does it detail specific train/test/validation splits for the other widely-used image classification datasets. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. It only mentions network backbones like ResNet-18. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch/TensorFlow, CUDA, specific libraries/packages). |
| Experiment Setup | Yes | For each model, we report maximum accuracy (over different combinations of learning rates based on different optimizers for θf and θg) based on Res Net-18 backbone in Table 1. The robustness is evaluated against PGD-10 attack (Kurakin, Goodfellow, and Bengio 2017) under ϵ = 0.5 and set step size = 1.25 (ϵ/step). The hyperparameter λ in this work is set to be 500. |