Collapse-Aware Triplet Decoupling for Adversarially Robust Image Retrieval
Authors: Qiwei Tian, Chenhao Lin, Zhengyu Zhao, Qian Li, Chao Shen
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on three datasets demonstrate that CA-TRIDE outperforms existing defense methods in both conventional and new metrics. |
| Researcher Affiliation | Academia | Qiwei Tian 1 Chenhao Lin 1 Zhengyu Zhao 1 Qian Li 1 Chao Shen 1 1Xi an Jiao Tong University, Xi an, China. Correspondence to: Chenhao Lin <linchenhao@xjtu.edu.cn>, Chao Shen <chaoshen@mail.xjtu.edu.cn>. |
| Pseudocode | Yes | Algorithm 1 Generating Adversarial Triplets in CA-TRIDE |
| Open Source Code | Yes | Codes are available at https://github.com/michaeltian108/CA-TRIDE. |
| Open Datasets | Yes | Evaluations are on three popular datasets in image retrieval tasks, i.e. CUB-200-2011 (Welinder et al., 2010), Cars196 (Krause et al., 2013), and SOP (Oh Song et al., 2016). |
| Dataset Splits | No | The paper describes a 'semi-hard sampling' strategy for mini-batch sampling and an epoch-wise adjustment to η, but it does not provide explicit train/validation/test dataset splits (e.g., percentages or sample counts) for reproduction. |
| Hardware Specification | Yes | We conducted 5 runs of HM and CA-TRIDE on the CUB dataset with an RTX3090 GPU. |
| Software Dependencies | No | The paper mentions using 'ADAM(Kingma & Ba, 2014) optimizer' and 'PGD (Madry et al., 2017)' but does not specify version numbers for these or any other software components or libraries used. |
| Experiment Setup | Yes | We train our models using ADAM(Kingma & Ba, 2014) optimizer with a 1.0 10 3 learning rate, a mini-batch size of 112, and training epochs of 100 under the above three datasets. For the top-rank pair, γ = 0.5 and the triplet margin in LT R βT R is 0.04. Adversarial perturbation is generated through PGD (Madry et al., 2017) with an optimization step α = 1/255, 16 iterations and clipped by an l norm of ϵ = 8/255. |