Vulnerabilities in Video Quality Assessment Models: The Challenge of Adversarial Attacks

Authors: Aoxiang Zhang, Yu Ran, Weixuan Tang, Yuan-Gen Wang

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this paper, we make the first attempt to evaluate the robustness of NR-VQA models against adversarial attacks, and propose a patch-based random search method for black-box attack. Extensive experiments have been conducted under both white-box and black-box settings, and the following conclusions can be drawn.
Researcher Affiliation Academia Ao-Xiang Zhang Yu Ran Weixuan Tang Yuan-Gen Wang Guangzhou University, China {zax, ranyu}@e.gzhu.edu.cn, {tweix, wangyg}@gzhu.edu.cn
Pseudocode Yes Algorithm 1 White-box attack on NR-VQA model. Algorithm 2 Black-box attack on NR-VQA model.
Open Source Code Yes The source code is available at https://github.com/GZHU-DVL/Attack VQA.
Open Datasets Yes The experiments are conducted on mainstream video datasets including Ko NVi D-1k [48], LIVE-VQC [15], You Tube UGC [45], and LSVQ [46].
Dataset Splits No The paper mentions using existing models and retraining them, and selecting 50 videos for evaluation, but it does not provide explicit training, validation, or test splits for the model training process itself.
Hardware Specification No The authors acknowledge the Network Center of Guangzhou University for providing HPC computing resources, but no specific hardware details such as GPU/CPU models, clock speeds, or memory configurations are provided.
Software Dependencies No The paper mentions 'Adam optimizer [47]' but does not provide specific software dependencies with version numbers (e.g., PyTorch 1.9, Python 3.8).
Experiment Setup Yes As for white-box attack, Adam optimizer [47] is applied, wherein the step size β is initialized as 3 10 4. K and T are set to 30 and 1. The pixel-level L2 norm of perturbations is constrained within 1/255. As for black-box attack, N, T, and γ are set to 300, 1, 5/255. Both h and w are set to 56.