Bias Removal Research Articles

The added value of diffusion tensor imaging with systematic bias correction for the assessment of liver morphology and physiology.

Diffusion-weighted images of the human liver are prone to artifacts from bulk motions, poor SNR, non-uniformity of magnetic field gradients, and non-optimal choice of diffusion weightings. These factors markedly affect diffusion tensor imaging (DTI) metrics such as mean diffusivity (MD) and fractional anisotropy (FA). This work presents a simple preprocessing pipeline for enhanced magnetic field gradient non-uniformity calibration and analysis of the systematic bias removal attained in each correction step. Liver DTI scans were conducted in two isotropic phantoms and one healthy volunteer. Diffusion tensor was calculated for the original data and after denoising, B1 correction, rigid body registration, and magnetic field gradient non-uniformity correction applying the B-matrix spatial distribution (BSD) method and then, compared with the standard approach (sDTI). MD and FA were determined in three segments of the right lobe from DTI using four different combinations of b-values from the set 0, 400, and 800 s/mm2. Results showed that the proposed preprocessing and BSD methods have a significant impact on MD and FA values in off- and iso-centered isotropic phantoms. The applied corrections applied to the human liver resulted in a 11% change in MD and a - 64% change in FA. By manipulating the b-values used in the diffusion tensor calculation, DTI metrics that reflect only morphology or additional information about liver tissue physiology can be obtained. Accurate quantification of the human liver by diffusion requires appropriate preprocessing and carefully chosen b-value. Noise, B1 inhomogeneity, mis-registration, and non-uniform magnetic field gradients significantly change distributions of DTI metrics in isotropic phantoms and the human liver. Basic preprocessing and the B-matrix spatial distribution (BSD) method perform differently for off-center and isocenter locations. In the human liver, they removed systematic bias of FA and MD by up to -63% and 11%, respectively. Visible variability of FA and MD among b-value sets indicates the possibility of DTI sensitization to different liver compartments.

Relative Homogenization of Climatic Time Series

Homogenization of the time series of observed climatic data aims to remove non-climatic biases caused by technical changes during the history of the climate observations. The spatial redundancy of climate information helps to recognize station-specific inhomogeneities with statistical methods, but the correct detection and removal of inhomogeneity biases is generally not easy for the combined effects of individual inhomogeneities. In a homogenization procedure, several time series of a given climatic variable observed in one climatic region are usually homogenized together via a large number of spatial comparisons between them. Such procedures are called relative homogenization. A relative homogenization procedure may include one or more homogenization cycles where a cycle includes the steps of time series comparison, inhomogeneity detection and corrections for inhomogeneities, and they may include other steps like the filtering of outlier values or spatial interpolations for infilling data gaps. Relative homogenization methods differ according to the number and content of the individual homogenization cycles, the procedure for the time series comparisons, the statistical inhomogeneity detection method, the way of the inhomogeneity bias removal, among other specifics. Efficient homogenization needs the use of tested statistical methods to be included in partly or fully automated homogenization procedures. Due to the large number and high variety of homogenization experiments fulfilled in the Spanish MULTITEST project (2015–2017), its method comparison test results are still the most informative about the efficiencies of homogenization methods in use. This study presents a brief review of the advances in relative homogenization, recalls some key results of the MULTITEST project, and analyzes some theoretical aspects of successful homogenization.

Compact Spatial Heterodyne Spectrographs for Future Space-Based Observations: Instrument Modeling and Applications.

High-resolution spectroscopy employing spatial heterodyne spectrographs (SHS) holds significant promise for forthcoming space missions, building upon its established track record in science applications. Notably, it offers exceptional performance and cost- effectiveness in the ultraviolet-visual (UV-Vis) region compared to contemporary instruments. SHS instruments provide high-resolution capabilities and substantially larger etendues than similar resolving power instruments. This study introduces a comprehensive Python-based SHS model integrated with a user-friendly web scraping interface for target star selection, parameter generation, and 2D interferogram creation. Our SHS model demonstrates double the resolving power of a grating spectrometer and a throughput comparable to a Fourier transform spectrometer (FTS) but without moving parts, enhancing robustness for deployment in space. The interferogram processing algorithm includes flat-fielding, bias removal, apodization, and an inverse Fourier transform (IFT) for accurate spectrum retrieval. Despite bandwidth limitations due to resolving power constraints, SHS models excel in applications requiring high spectral resolution over narrow wavelength ranges, such as studying isotopic emission lines. The model provides optimization results and trade-offs for system parameters, ensuring precise spectral recovery with realistic signal-to-noise ratio (SNR) values. SHS is versatile and effective for various scientific applications, including investigating atomic and molecular emissions from comets, planetary atmospheres, the Earth's atmosphere, the Sun, and the interstellar medium (ISM). This research significantly contributes to expediting the development and deployment of SHS instruments, demonstrating their potential across numerous scientific domains.

Mitigating implicit and explicit bias in structured data without sacrificing accuracy in pattern classification

AbstractUsing biased data to train Artificial Intelligence (AI) algorithms will lead to biased decisions, discriminating against certain groups or individuals. Bias can be explicit (one or several protected features directly influence the decisions) or implicit (one or several protected features indirectly influence the decisions). Unsurprisingly, biased patterns are difficult to detect and mitigate. This paper investigates the extent to which explicit and implicit against one or more protected features in structured classification data sets can be mitigated simultaneously while retaining the data’s discriminatory power. The main contribution of this paper concerns an optimization-based bias mitigation method that reweights the training instances. The algorithm operates with numerical and nominal data and can mitigate implicit and explicit bias against several protected features simultaneously. The trade-off between bias mitigation and accuracy loss can be controlled using parameters in the objective function. The numerical simulations using real-world data sets show a reduction of up to 77% of implicit bias and a complete removal of explicit bias against protected features at no cost of accuracy of a wrapper classifier trained on the data. Overall, the proposed method outperforms the state-of-the-art bias mitigation methods for the selected data sets.

Removal of historical taxonomic bias and its impact on biogeographic analyses: a case study of Neotropical tardigrade fauna

Abstract Large-scale databases are crucial for macroecology research, yet not entirely bias-free. Studying the biogeography of metazoan microfauna utilizing such databases is challenging, and added biases can further hinder them. Incidence data of tardigrade species from Central and South America are geographically biased, comprising mostly species erroneously considered cosmopolitan. The impact of these biases on macroecological studies is still unknown. This paper evaluates how biased distribution records affect understanding of non-marine tardigrade distribution patterns in the Neotropical region. Using two datasets (including and excluding dubious records of allegedly cosmopolitan species), we assess how well biogeographic regions are sampled, estimate tardigrade species’ richness, and analyse whether including dubious records can alter our comprehension of their macroecology. We demonstrate how biogeographic regions are still not fully sampled, and that including unreliable records influences how many species remain to be discovered. Since records of ‘false cosmopolitan species’ represent most records, their removal increases uncertainty while yielding distribution patterns more likely to be accurate. Disregarding untrustworthy distribution records of ‘false cosmopolitan species’ is the first step to better understanding tardigrade macroecology in the Neotropics, and probably worldwide. However, additional, preferably systematic sampling is required before we can infer general tardigrade biodiversity patterns in under-sampled regions.

Parkinson’s Disease Recognition Using Decorrelated Convolutional Neural Networks: Addressing Imbalance and Scanner Bias in rs-fMRI Data

Parkinson’s disease (PD) is a neurodegenerative and progressive disease that impacts the nerve cells in the brain and varies from person to person. The exact cause of PD is still unknown, and the diagnosis of PD does not include a specific objective test with certainty. Although deep learning has made great progress in medical neuroimaging analysis, these methods are very susceptible to biases present in neuroimaging datasets. An innovative decorrelated deep learning technique is introduced to mitigate class bias and scanner bias while simultaneously focusing on finding distinguishing characteristics in resting-state functional MRI (rs-fMRI) data, which assists in recognizing PD with good accuracy. The decorrelation function reduces the nonlinear correlation between features and bias in order to learn bias-invariant features. The publicly available Parkinson’s Progression Markers Initiative (PPMI) dataset, referred to as a single-scanner imbalanced dataset in this study, was used to validate our method. The imbalanced dataset problem affects the performance of the deep learning framework by overfitting to the majority class. To resolve this problem, we propose a new decorrelated convolutional neural network (DcCNN) framework by applying decorrelation-based optimization to convolutional neural networks (CNNs). An analysis of evaluation metrics comparisons shows that integrating the decorrelation function boosts the performance of PD recognition by removing class bias. Specifically, our DcCNN models perform significantly better than existing traditional approaches to tackle the imbalance problem. Finally, the same framework can be extended to create scanner-invariant features without significantly impacting the performance of a model. The obtained dataset is a multiscanner dataset, which leads to scanner bias due to the differences in acquisition protocols and scanners. The multiscanner dataset is a combination of two publicly available datasets, namely, PPMI and FTLDNI—the frontotemporal lobar degeneration neuroimaging initiative (NIFD) dataset. The results of t-distributed stochastic neighbor embedding (t-SNE) and scanner classification accuracy of our proposed feature extraction–DcCNN (FE-DcCNN) model validated the effective removal of scanner bias. Our method achieves an average accuracy of 77.80% on a multiscanner dataset for differentiating PD from a healthy control, which is superior to the DcCNN model trained on a single-scanner imbalanced dataset.

Optimal Fitting and Debiasing for Detectors Read Out Up-the-Ramp

This paper derives the optimal fit to a pixel’s count rate in the case of an ideal detector read out nondestructively in the presence of both read and photon noise. The approach is general for any readout scheme, provides closed-form expressions for all quantities, and has a computational cost that is linear in the number of resultants (groups of reads). I also derive the bias of the fit from estimating the covariance matrix and show how to remove it to first order. The ramp-fitting algorithm I describe provides the χ 2 value of the fit of a line to the accumulated counts, which can be interpreted as a goodness-of-fit metric. I provide and describe a pure Python implementation of these algorithms that can process a 10-resultant ramp on a 4096 × 4096 detector in ≈8 s with bias removal on a single core of a 2020 Macbook Air. This Python implementation, together with tests and a tutorial notebook, are available at https://github.com/t-brandt/fitramp. A companion paper describes a jump detection algorithm based on hypothesis testing of ramp fits and demonstrates all algorithms on data from JWST.

Single‐shot phase‐shifting on Michelson interferometry for incoherent digital holography

AbstractA single‐shot phase‐shifting method in incoherent digital holography (IDH) based on Michelson interferometry is proposed herein. The proposed method uses polarization‐modulating optical elements and a polarization sensor in the Michelson interferometer optics to induce a geometric phase shift. It acquires four holograms with different geometric phase retardations in a single exposure to eliminate the bias and conjugate noise using a phase‐shifting technique. The proposed optical system enables real‐time bias and conjugate term removal from a recorded hologram without requiring sequential recording for phase shifting. Furthermore, because it is based on Michelson interferometry, it can be implemented using simple reflective optics without expensive or difficult‐to‐produce optical elements for single‐shot phase shifting. The principles of self‐interference and geometric phase shifting in the proposed optical structure were analyzed mathematically. A real‐time IDH‐recording experimental setup was constructed to verify the proposed method. The experimental results of single‐shot phase‐shifting IDH and hologram movie recordings confirm the feasibility of the proposed method.

Robust Underwater Acoustic Channel Estimation Method Based on Bias-Free Convolutional Neural Network

In recent years, the study of deep learning techniques for underwater acoustic channel estimation has gained widespread attention. However, existing neural network channel estimation methods often overfit to training dataset noise levels, leading to diminished performance when confronted with new noise levels. In this research, a “bias-free” denoising convolutional neural network (DnCNN) method is proposed for robust underwater acoustic channel estimation. The paper offers a theoretical justification for bias removal and customizes the fundamental DnCNN framework to give a specialized design for channel estimation, referred to as the bias-free complex DnCNN (BF-CDN). It uses least squares channel estimation results as input and employs a CNN model to learn channel characteristics and noise distribution. The proposed method effectively utilizes the temporal correlation inherent in underwater acoustic channels to further enhance estimation performance and robustness. This method adapts to varying noise levels in underwater environments. Experimental results show the robustness of the method under different noise conditions, indicating its potential to improve the accuracy and reliability of channel estimation.

Comparison of postoperative hypersensitivity between Total-etch and Universal adhesive system: a randomized clinical trial

The objective was to determine the postoperative hypersensitivity of two-step Total-etch as compared to one-step Universal adhesives followed by composite restorations on 100 patients by applying Total-etch on one tooth and Universal adhesive on another tooth. The bonds and teeth were randomly selected. Postoperative hypersensitivity was recorded by visual analog scale before, immediately after, and 24 h after the restoration using cold stimulus. The Mann–Whitney test was applied for statistical comparison of postoperative hypersensitivity between the two bonds as well as for any significant difference in genders with each bond. No significant difference was found between postoperative hypersensitivity of the two adhesives before (p-value = 0.57), immediately after (p-value = 0.604), and 24 h after (p-value = 0.728) the restoration. Males showed more hypersensitivity with Total-etch as compared to females before (p-value = 0.037), immediately after (p-value = 0.047), and 24 h after the restoration (p-value = 0.022). No significant difference was found between gender and Universal adhesive at all three stages (p-value > 0.05). The results suggest no significant difference in postoperative hypersensitivity between the two materials when good sample size and proper technique were observed along with the removal of bias like different patients having different pain perceptions and multiple operators having different operating skills. Males showed more hypersensitivity to Total-etch.Trial registration number: Australian New Zealand Clinical Trials. Registry number: ACTRN12622001213730. (Retrospectively registered: 09/09/2022).

Peer review in the global digital age: perspectives of publishing industry stakeholders

Peer review is a crucial component of the scientific publication process, enabling validation of research, identification of errors, and removal of potential bias. However, there are some well-known limitations, including slow publication cycles and overstringent gatekeeping. Artificial intelligence and digital technology are revolutionizing peer review and publishing by addressing some of the limitations, and fostering closer collaboration among scholars worldwide.1-3 This paradigm shift aligns with the principles of open science, enhancing the reach and impact of scholarly work. Digital tools for peer review are already transforming many aspects of this process, by enhancing quality control, automation of routine tasks, and expediting laborious aspects of the peer review process, thereby enhancing speed and efficiency. Digital platforms are reducing publication times and potentially allowing for the promotion of diversity and inclusivity of the peer reviewer pool by vastly enhancing global connectivity. Selecting qualified and impartial global reviewers in the digital context is vital for the future of our rapidly evolving and increasingly diverse publication landscape. Editors play a key role in oversight while providing reviewers with clear guidelines and training. In conclusion, digital tools assisting peer review will inevitably play an increasingly useful role in enhancing the efficiency, and potentially the inclusivity and objectivity of the process.

An attention-mechanism-based deep fusion model for improving quantitative precipitation estimation in a sparsely-gauged basin

Improving quantitative precipitation estimation (QPE) in sparsely-gauged basins via merging remote sensing precipitation data and rain gauge data still remains a challenge since most existing merging models degrade when the rain gauge data become limited. To address the challenge, we propose an attention-mechanism-based deep learning model, Multi-Level Transformer Fusion (MLTF) model, which allows to capture the inner interactions among the multi-source input data (TRMM 3B42 V7 data, GridSat-B1 data, and DEM) in the sparsely-gauged basin. Taking the source region of the Yellow River basin (SRYRB) as a representative case study, we demonstrate the performance of the proposed model and compare it with conventional methods (e.g., Multiplicative Bias Correction, Additive Bias Removal, linear regression, and Kriging) and deep-learning-based models (CNN, CNN-LSTM). Results indicate that the merged precipitation in SRYRB produced by the MLTF model exhibits a RMSE reduction of 27.1 %, MAE decrease of 11.2 %, and CC increase of 19.2 % in comparison to the original TRMM data, outperforming all the selected comparative methods. Finally, an improved daily precipitation dataset during 1999–2019 with a spatial resolution of 0.05° is produced for the study area. This study proposes a new method for QPE improvement in a sparsely-gauged basin, which would provide valuable data support for regional hydrological study and water resources management.

A Comprehensive Step-by-Step Guide to Using Data Science Tools in the Gestion of Epidemiological and Climatological Data in Rice Production Systems

The application of data science (DS) techniques has become increasingly essential in various fields, including epidemiology and climatology in agricultural production systems. In this sector, traditionally large amounts of data are acquired, but not well-managed and -analyzed as a basis for evidence-based decision-making processes. Here, we present a comprehensive step-by-step guide that explores the use of DS in managing epidemiological and climatological data within rice production systems under tropical conditions. Our work focuses on using the multi-temporal dataset associated with the monitoring of diseases and climate variables in rice in Colombia during eight years (2012–2019). The study comprises four main phases: (I) data cleaning and organization to ensure the integrity and consistency of the dataset; (II) data management involving web-scraping techniques to acquire climate information from free databases, like WordClim and Chelsa, validation against in situ weather stations, and bias removal to enrich the dataset; (III) data visualization techniques to effectively represent the gathered information, and (IV) a basic analysis related to the clustering and climatic characterization of rice-producing areas in Colombia. In our work, a process of evaluation and the validation of climate data are conducted based on errors (r, R2, MAE, RSME) and bias evaluation metrics. In addition, in phase II, climate clustering was conducted based on a PCA and K-means algorithm. Understanding the association of climatic and epidemiological data is pivotal in predicting and mitigating disease outbreaks in rice production areas. Our research underscores the significance of DS in managing epidemiological and climatological data for rice production systems. By applying a protocol responsible for DS tools, our study provides a solid foundation for further research into disease dynamics and climate interactions in rice-producing regions and other crops, ultimately contributing to more informed decision-making processes in agriculture.

Development of correlation ECE system for electron temperature fluctuation measurement in LHD

Correlation-ECE (C-ECE) is a standard method for investigating turbulence driven transport. This method allows electron temperature fluctuations that contain information on turbulent transport and independent thermal noise. The turbulence feature is extracted from a correlation analysis from two close locations. The A C-ECE system is utilized on the large helical device (LHD) to measure emission within the frequency range of 74–79.6 GHz. This system employs the spectral decorrelation method and serves as a collective Thomson scattering diagnostic receiver in the LHD. The C-ECE receiver system is comprised of a filter bank system with 32 band-pass filters and a fast digitizer system operating at a sampling rate of 12.5 GHz in the intermediate frequency (IF) stage. This study presents initial experimental results on temperature fluctuation spectra in the LHD, obtained through the C-ECE system using a coherency-based analysis method. An MHD mode at 5 kHz is excited from the onset of neutral beam injection in a magnetic probe, and coherence spectra are obtained from two C-ECE receiver systems. The temperature fluctuation results are derived from the coherence spectrum after bias removal and indicate a level of approximately 3% in the frequency range of 0 to 400 kHz. Further investigations will be conducted to explore drift wave turbulence activities and reconstruct the radial profile of temperature fluctuation in the LHD using the C-ECE receiver systems.

Functional diversity metrics can perform well with highly incomplete data sets

Abstract Characterising changes in functional diversity at large spatial scales provides insight into the impact of human activity on ecosystem structure and function. However, the approach is often based on trait data sets that are incomplete and unrepresentative, with uncertain impacts on functional diversity estimates. To address this knowledge gap, we simulated random and biased removal of data from three empirical trait data sets: an avian data set (9579 species), a plant data set (2185 species) and a crocodilian data set (25 species). For these data sets, we assessed whether functional diversity metrics were robust to data incompleteness with and without using imputation to fill data gaps. We compared two metrics each calculated with two methods: functional richness (calculated with convex hulls and trait probabilities densities) and functional divergence (calculated with distance‐based Rao and trait probability densities). Without imputation, estimates of functional diversity (richness and divergence) for birds and plants were robust when 20%–70% of species had missing data for four out of 11 and two out of six continuous traits, respectively, depending on the severity of bias and method used. However, when missing traits were imputed, functional diversity metrics consistently remained representative of the true value when 70% of bird species were missing data for four out of 11 traits and when 50% of plant species were missing data for two out of six traits. Trait probability densities and distance‐based Rao were particularly robust to missingness and bias when combined with imputation. Convex hull‐based estimations of functional richness were less reliable. When applied to a smaller data set (crocodilians, 25 species), all functional diversity metrics were much more sensitive to missing data. Expanding global morphometric data sets to represent more taxa and traits, and to quantify intraspecific variation, remains a priority. In the meantime, our results show that widely used methods can successfully quantify large‐scale functional diversity even when data are missing for half of species, provided that missing traits are estimated using imputation. We recommend the use of trait probability densities or distance‐based Rao when working with large incomplete data sets and filling data gaps with imputation.

Modelling extreme water levels using intertidal topography and bathymetry derived from multispectral satellite images

Abstract. Topographic and bathymetric data are essential for accurate predictions of flooding in estuaries because water depth and elevation data are fundamental components of the shallow-water hydrodynamic equations used in models for storm surges and tides. Where lidar or in situ acoustic surveys are unavailable, recent efforts have centred on using satellite-derived bathymetry (SDB) and satellite-derived topography (SDT). This work is aimed at (1) determining the accuracy of SDT and (2) assessing the suitability of the SDT and SDB for extreme water level modelling of estuaries. The SDT was created by extracting the waterline as it tracks over the topography with changing tides. The method was applied to four different estuaries in Aotearoa / New Zealand: Whitianga, Maketū, Ōhiwa and Tauranga harbours. Results show that the waterline method provides similar topography to the lidar with a root-mean-square error equal to 0.2 m, and it is slightly improved when two correction methods are applied to the topography derivations: the removal of statistical bias (0.02 m improvement) and hydrodynamic modelling correction of waterline elevation (0.01 m improvement). The use of SDT in numerical simulations of surge levels was assessed for Tauranga Harbour in eight different simulation scenarios. Each scenario explored different ways of incorporating the SDT to replace the topographic data collected using non-satellite survey methods. In addition, one of these scenarios combined SDT (for intertidal zones) and SDB (for subtidal bathymetry), so only satellite information is used in surge modelling. The latter SDB is derived using the well-known ratio–log method. For Tauranga Harbour, using SDT and SDB in hydrodynamic models does not result in significant differences in predicting high water levels when compared with the scenario modelled using surveyed bathymetry.

Deep Learning-Based Detection of Bone Tumors around the Knee in X-rays of Children.

Even though tumors in children are rare, they cause the second most deaths under the age of 18 years. More often than in other age groups, underage patients suffer from malignancies of the bones, and these mostly occur in the area around the knee. One problem in the treatment is the early detection of bone tumors, especially on X-rays. The rarity and non-specific clinical symptoms further prolong the time to diagnosis. Nevertheless, an early diagnosis is crucial and can facilitate the treatment and therefore improve the prognosis of affected children. A new approach to evaluating X-ray images using artificial intelligence may facilitate the detection of suspicious lesions and, hence, accelerate the referral to a specialized center. We implemented a Vision Transformer model for image classification of healthy and pathological X-rays. To tackle the limited amount of data, we used a pretrained model and implemented extensive data augmentation. Discrete parameters were described by incidence and percentage ratio and continuous parameters by median, standard deviation and variance. For the evaluation of the model accuracy, sensitivity and specificity were computed. The two-entity classification of the healthy control group and the pathological group resulted in a cross-validated accuracy of 89.1%, a sensitivity of 82.2% and a specificity of 93.2% for test groups. Grad-CAMs were created to ensure the plausibility of the predictions. The proposed approach, using state-of-the-art deep learning methodology to detect bone tumors on knee X-rays of children has achieved very good results. With further improvement of the algorithm, enlargement of the dataset and removal of potential biases, this could become a useful additional tool, especially to support general practitioners for early, accurate and specific diagnosis of bone lesions in young patients.

Innovative remote sensing methodologies and applications in coastal and marine environments

ABSTRACT Remote sensing (RS) technologies are extensively exploited by scientists and a vast audience of local authorities, urban managers, and city planners. Coastal regions, geohazard-prone areas, and highly populated cities represent natural laboratories to apply RS technologies and test new methods. Over the last decades, many efforts have been spent on improving Earth’s surface monitoring, including intensifying Earth Observation (EO) operations by the major national space agencies. They oversee to plan and make operational constellations of satellite sensors providing the scientific community with extensive research and development opportunities in the geoscience field. For instance, within this framework, the European Space Agency (ESA) and the Ministry of Science and Technology of China (MOST) have sponsored, since the early 2000s, the DRAGON initiative jointly carried out by the European and Chinese RS scientific communities. This manuscript aims to provide a synthetic overview of some research activities and new methods recently designed and applied and trace the route for further developments. The main findings are related to i) the analysis of flood risk in China, ii) the potential of new methods for the estimation and removal of ground displacement biases in small-baseline oriented interferometric Synthetic Aperture Radar (SAR) methods, iii) the analysis of the inundation risk in low-lying regions using coherent and incoherent SAR methods; and iv) the use of SAR-based technologies for marine applications.

Microphone Based Acoustic Vector Sensor for Direction Finding with Bias Removal

Microphone Based Acoustic Vector Sensor for Direction Finding with Bias Removal

Revisiting model fairness via adversarial examples

Existing research literally evaluates model fairness over limited observed data. In practice, however, factors such as maliciously crafted examples and naturally corrupted examples often appear in real-world data collection. This severely limits the reliability of bias removal methods, inhibits fairness improvement in long-term learning systems, and probes to study accuracy-related robustness. Therefore, we ask: How adversarial examples will skew model fairness? In this paper, we investigate the vulnerability of individual fairness and group fairness to adversarial attacks. We further propose a general adversarial fairness attack framework capable of twisting model bias through a small subset of adversarial examples. We formulate this problem as an optimization problem: maximizing the model bias with the constraint of the number of adversarial examples and the perturbation scale. Our approach finds the most vulnerable examples to model fairness based on the estimated distance from examples to the decision boundary and demographic information. The experimental results 11https://github.com/TaocsZhang/Fairness-Attack-via-Adversarial-Examples. show that model fairness is easily skewed by a small number of adversarial examples.