Perceptual Attacks of No-Reference Image Quality Models with Human-in-the-Loop

Authors: Weixia Zhang, Dingquan Li, Xiongkuo Min, Guangtao Zhai, Guodong Guo, Xiaokang Yang, Kede Ma

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

Reproducibility Variable Result LLM Response
Research Type Experimental We test one knowledge-driven and three data-driven NR-IQA methods under four full-reference IQA models (as approximations to human perception of just-noticeable differences). Through carefully designed psychophysical experiments, we find that all four NR-IQA models are vulnerable to the proposed perceptual attack. More interestingly, we observe that the generated counterexamples are not transferable, manifesting themselves as distinct design flows of respective NR-IQA methods. Source code are available at https://github.com/zwx8981/Perceptual Attack_BIQA. We conduct an extensive experiment to examine four NR-IQA models, the knowledge-driven BRISQUE [17], the shallow learning-based CORNIA [21], as well as the deep learningbased Ma19 [22] and UNIQUE [6] under four FR-IQA models, the Chebyshev distance (i.e., the ℓ -norm induced metric), SSIM, LPIPS, and DISTS (as approximations to human perception of JNDs).
Researcher Affiliation Academia 1 Mo E Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 Network Intelligence Research Department, Peng Cheng Laboratory 3 Department of Computer Science and Electrical Engineering, West Virginia University 4 Department of Computer Science, City University of Hong Kong {zwx8981, minxiongkuo, zhaiguangtao, xkyang}@sjtu.edu.cn, lidq01@pcl.ac.cn guodong.guo@mail.wvu.edu, kede.ma@cityu.edu.hk
Pseudocode Yes Algorithm 1 Perceptually Imperceptible Counterexample Generation
Open Source Code Yes Source code are available at https://github.com/zwx8981/Perceptual Attack_BIQA.
Open Datasets Yes We use the training codes provided by the original authors to re-train BRISQUE and CORNIA on LIVE [7], Ma19 [22] on our own collected dataset4, and UNIQUE on six human-rated IQA databases [7, 63, 64, 9, 5, 65]. We collect twelve images as initializations from the publicly available LIVE IQA database [7] (see Fig. 3)...
Dataset Splits No The paper states which datasets were used for training different models (e.g., "re-train BRISQUE and CORNIA on LIVE [7]"), but it does not specify the train/validation/test splits (e.g., percentages or sample counts) used for this training or for their own experiments beyond selecting "twelve initial images" from LIVE.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU models, CPU types) used for running the experiments.
Software Dependencies No The paper does not provide specific software dependency details with version numbers (e.g., "Python 3.8, PyTorch 1.9"). It mentions "training codes provided by the original authors" but no versions.
Experiment Setup Yes For each of sixteen combinations of NR-IQA and FR-IQA models, and each of the twelve initial images, we set λ to 32 values, and optimize the objective in Eq. (2) to generate 32 perturbed images... We set the step size γ to 10 3 and the maximum number of iterations to 200, respectively. As suggested by the BT. 500 recommendations [69], we carry out the experiments in an indoor office environment with a normal lighting condition (with approximately 200 lux) and without reflecting ceiling walls and floors. The peak luminance of the displayed images is mapped to 200 cd/m2. We recruit fifteen human subjects (with normal or corrected-to-normal vision) to participate in the psychophysical experiment, viewing the image pairs from a fixed distance of twice the screen height.