Few-Shot Backdoor Attacks on Visual Object Tracking

Authors: Yiming Li, Haoxiang Zhong, Xingjun Ma, Yong Jiang, Shu-Tao Xia

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We examine our attack in both digital and physical-world settings and show that it can significantly degrade the performance of state-of-the-art VOT trackers. We also show that our attack is resistant to potential defenses, highlighting the vulnerability of VOT models to potential backdoor attacks. ... 4 EXPERIMENTS
Researcher Affiliation Academia Yiming Li1, Haoxiang Zhong1,2, Xingjun Ma3, Yong Jiang1,2, Shu-Tao Xia1,2 1Tsinghua Shenzhen International Graduate School, Tsinghua University, China 2Research Center of Artificial Intelligence, Peng Cheng Laboratory, China 3School of Computer Science, Fudan University, China {li-ym18, zhx19}@mails.tsinghua.edu.cn; danxjma@gmail.com; {jiangy, xiast}@sz.tsinghua.edu.cn
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes The codes for reproducing the main experiments of our FSBA are also open-sourced, as described in Appendix L. ... The codes for reproducing the main experiments of our FSBA are open-sourced on Github7. 7https://github.com/HXZhong1997/FSBA
Open Datasets Yes We evaluate the effectiveness of BOBA and our FSBA attack on three advanced siamese network based trackers... on OTB100 (Wu et al., 2015) and GOT10K (Huang et al., 2019) datasets. ... We also provide the results on the La SOT dataset (Fan et al., 2019) in Appendix C. ... we train the Siam RPN++ with a backbone of Res Net-50 (He et al., 2016) only on COCO (Lin et al., 2014), ILSVRC-DET (Russakovsky et al., 2015), and ILSVRC-VID (Russakovsky et al., 2015) datasets...
Dataset Splits Yes The GOT10K (Huang et al., 2019) is a large and highly diverse dataset. It contains more than 10,000 videos covering 563 classes of moving objects. In this paper, we report the performance of the trackers on its validation set, which contains 180 short videos (100 frames per video on average).
Hardware Specification Yes on a single NVIDIA 2080Ti; ... with four NVIDIA V100 GPUs
Software Dependencies No The paper mentions using 'Pytorch (Paszke et al., 2019)' and 'torchvision' but does not provide specific version numbers for these software components.
Experiment Setup Yes Specifically, for the benign model, we use the SGD optimizer with momentum 0.9, weight decay of 5 10 4, and an initial learning rate of 0.01. An exponential learning rate scheduler is adopted with a final learning rate of 10 5. We train the model for 50 epochs with a batch size of 8...