Attacks on Continual Semantic Segmentation by Perturbing Incremental Samples

Authors: Zhidong Yu, Wei Yang, Xike Xie, Zhenbo Shi

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that on the popular datasets Pascal VOC, ADE20k, and Cityscapes, our approach easily degrades the performance of currently popular CSS methods, which reveals the importance of security in CSS. We conduct extensive experiments at multiple incremental settings on the standard benchmarks, and the proposed method substantially reduces the performance of CSS on old and new classes.
Researcher Affiliation Academia 1School of Computer Science and Technology, University of Science and Technology of China, Hefei 230026, China 2Hefei National Laboratory, Hefei 230088, China 3Suzhou Institute for Advanced Research, University of Science and Technology of China, Suzhou 215123, China
Pseudocode Yes Algorithm 1: Attack process at step t Input: Segmentation model Mt 1 and incremental data Dt Parameter: Attack model Gt and segmentation model Mt at step t Output: Segmentation model Mt 1: while Current epochs less than total epochs do 2: while {(Xt, Y t)} Dt do 3: Get the Perturbations: r = G(Xt) 4: Get the perturbed image with Eq. (2) 5: Get original and perturbed predictions by Mt 1. 6: Update Gt with Eq. (7) 7: end while 8: end while 9: Generate the Perturbations r for Xt Dt and construct Adversarial Data ˆDt 10: Incremental Training for Mt with ˆDt 11: return Mt
Open Source Code No The paper does not provide any explicit statement or link indicating that the source code for the methodology described is open-source or publicly available.
Open Datasets Yes Datasets. We validate CS-Attack on different standard semantic segmentation datasets: Pascal VOC2012 (Everingham et al. 2010), ADE20k (Zhou et al. 2017) and Cityscapes (Cordts et al. 2016).
Dataset Splits Yes The Pascal VOC2012 dataset contains 20 object classes and the background. It includes 10,582 images for training and 1,449 images for validation, respectively. The Cityscapes dataset contains 19 classes from 21 cities with 2,975 training images, 500 validation images and 1,525 test images.
Hardware Specification Yes Our experiments are conducted on 4 NVIDIA 2080Ti GPUs.
Software Dependencies No The paper mentions the use of Deeplabv3 architecture but does not specify software dependencies with version numbers (e.g., Python version, specific deep learning framework versions like PyTorch or TensorFlow, or other library versions).
Experiment Setup Yes To train the model Gt, we use the stochastic gradient descent (SGD) optimizer, where the base learning rate is 0.01 for all datasets, and λ is 1 in our experiments. Gt is trained for 30 epochs on PASCAL VOC2012 and ADE20k datasets and 50 epochs on Cityscapes dataset.