Perceptual Fairness in Image Restoration

Authors: Guy Ohayon, Michael Elad, Tomer Michaeli

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the superiority of PF over RDP in detecting fairness bias in face image superresolution. Our analysis considers various aspects, including different types of degradations, and fairness evaluations across four groups categorized by ethnicity and age. First, we show that RDP incorrectly attributes fairness in a simple scenario where fairness is clearly violated. In contrast, PF successfully detects the bias. Second, we showcase a scenario where PF uncovers potential malicious intent. Specifically, it can detect bias injected into the system via adversarial attacks, a situation again missed by RDP.
Researcher Affiliation Academia Guy Ohayon Faculty of Computer Science Technion Israel Institute of Technology ohayonguy@cs.technion.ac.il Michael Elad Faculty of Computer Science Technion Israel Institute of Technology elad@cs.technion.ac.il Tomer Michaeli Faculty of Electrical and Computer Engineering Technion Israel Institute of Technology tomer.m@ee.technion.ac.il
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No We do not release data or models. The paper poses no such risks.
Open Datasets Yes Such methods are often trained and evaluated on high-quality, aligned face image datasets like Celeb A-HQ [36] and FFHQ [37], which lack ground truth labels for sensitive attributes such as ethnicity.
Dataset Splits No The paper mentions evaluating on a "test partition" of Celeb A-HQ but does not specify a separate validation or training split percentage or methodology for its own experiments.
Hardware Specification Yes All our experiments are conducted on a NVIDIA RTX A6000 GPU.
Software Dependencies Yes We use the torch-fidelity package [57] (Git Hub commit a61422f) to compute the KID [9], FID [30], precision and recall [45]. The GPSNR and the GLPIPS are computed using the piq package [38, 39] (version 0.8.0 in pip).
Experiment Setup Yes All algorithms are evaluated using the official codes and checkpoints provided by their authors. We do not optimize these algorithms within this work. However, we do conduct adversarial attacks, which require optimization. We disclose the hyper-parameters used in such experiments.