Association for the Advancement of Artificial Intelligence: AAAI Publications
Not a member yet
26155 research outputs found
Sort by
Exploring Disparity-Accuracy Trade-offs in Face Recognition Systems: The Role of Datasets, Architectures, and Loss Functions
No full textAutomated Face Recognition Systems (FRSs), developed using deep learning models, are deployed worldwide for identity verification and facial attribute analysis. The performance of these models is determined by a complex interdependence among the model architecture, optimization/loss function and datasets. Although FRSs have surpassed human-level accuracy, they continue to be disparate against certain demographics. Due to the ubiquity of applications, it is extremely important to understand the impact of the three components-- model architecture, loss function and face image dataset on the accuracy-disparity trade-off to design better, unbiased platforms. In this work, we perform an in-depth analysis of three FRSs for the task of gender prediction, with various architectural modifications resulting in ten deep-learning models coupled with four loss functions and benchmark them on seven face datasets across 266 evaluation configurations. Our results show that all three components have an individual as well as a combined impact on both accuracy and disparity. We identify that datasets have an inherent property that causes them to perform similarly across models, independent of the choice of loss functions. Moreover, the choice of dataset determines the model's perceived bias-- the same model reports bias in opposite directions for three gender-balanced datasets of ``in-the-wild'' face images of popular individuals. The facial embeddings show that the models are unable to generalize a uniform definition of what constitutes a ``female face'' as opposed to a ``male face'', due to dataset diversity. We provide recommendations to model developers on using our study as a blueprint for development and subsequent deployment
Algorithmic Behaviors Across Regions: A Geolocation Audit of YouTube Search for COVID-19 Misinformation Between the United States and South Africa
No full textDespite being an integral tool for finding health-related information online, YouTube has faced criticism for disseminating COVID-19 misinformation globally to its users. Yet, prior audit studies have predominantly investigated YouTube within the Global North contexts, often overlooking the Global South. To address this gap, we conducted a comprehensive 10-day geolocation-based audit on YouTube to compare the prevalence of COVID-19 misinformation in search results between the United States (US) and South Africa (SA), the countries heavily affected by the pandemic in the Global North and the Global South, respectively. For each country, we selected 3 geolocations and placed sock-puppets, or bots emulating "real" users, that collected search results for 48 search queries sorted by 4 search filters for 10 days, yielding a dataset of 915K results. We found that 31.55% of the top-10 search results contained COVID-19 misinformation. Among the top-10 search results, bots in SA faced significantly more misinformative search results than their US counterparts. Overall, our study highlights the contrasting algorithmic behaviors of YouTube search between two countries, underscoring the need for the platform to regulate algorithmic behavior consistently across different regions of the Globe.
Warning: We caution the readers that some examples provided to better contextualize our data can be offensive
Exploiting Explainability to Design Adversarial Attacks and Evaluate Attack Resilience in Hate-Speech Detection Models
No full textThe advent of social media has given rise to numerous ethical challenges, with hate speech among the most significant concerns. Researchers are attempting to tackle this problem by using hate-speech detection and employing language models to automatically moderate content and promote civil discourse. Unfortunately, recent studies have revealed that hate-speech detection systems can be misled by adversarial attacks, raising concerns about their resilience. While previous research has separately addressed the robustness of these models under adversarial attacks and their explainability, there has been no comprehensive study exploring their intersection. The novelty of our work lies in combining these two critical aspects, leveraging explainability to identify potential vulnerabilities and enabling the design of targeted adversarial attacks. This paper quantifies the interplay between explainability and adversarial robustness in hate-speech detection models. We define novel metrics based on explainability-driven adversarial attacks to evaluate this relationship, providing a clear assessment of model vulnerabilities and guiding the development of more resilient systems
Online Myths on Opioid Use Disorder: A Comparison of Reddit and Large Language Model
No full textOnline communities on Reddit are a popular choice among people with opioid use disorder (OUD) to seek information on drug use, withdrawal symptoms, and recovery. LLM-powered chatbots (e.g., ChatGPT) are widely being adopted as question-answer systems for health-related queries. However, such online health information seeking could potentially be hindered by myths and misinformation on OUD, misleading or causing genuine harm to people with OUD. In this work, we examine the prevalence of 5 OUD-related myths, on treatment models and patient characteristics, within human- (taken from Reddit) and LLM-generated responses to queries on OUD. We further explore the framing strategies used within responses (both human- and LLM-generated) promoting and countering the myths. We found that all 5 myths were more widespread within human-generated responses. In addition, myth-promoting responses adopted trustworthy and authoritative framings, compared to knowledge-imparting linguistic cues within those countering the myths. Our work offers recommendations to reduce online OUD misinformation
Towards Characterizing and Detecting Incentivized Reviews on eCommerce Platforms
No full textCustomer reviews play an important role in rankings and visibility on e-commerce sites, and also strongly influence a customer's decision to purchase a product. Motivated by this, malicious sellers engage in incentivized review fraud to inflate their product ratings by providing customers with free products in exchange for five-star reviews, thus compromising review integrity. While there is ample prior work on fake reviews in general, there is limited prior work on incentivized review fraud. In this work, we infiltrate an underground market for fake reviews and implement a custom crawler to collect a dataset of malicious products that seek incentivized reviews. We devise and extract a set of features, and show that these are statistically significant in differentiating between benign and malicious products. Using hypothesis testing, we identify characteristics and trends exhibited by malicious products. While we are unable to achieve a high precision when we train standard machine learning models without compromising on the recall, we propose a lightweight two-phase technique that combines high-precision product classifiers with high-recall review classifiers. This technique allows us to minimize the false positives, with only a slight increase in false negatives. Finally, we also audit the effectiveness of two publicly available tools for incentivized review detection and find that they are not reliable. In summary, we contribute a new high-fidelity dataset, characterize products seeking incentivized reviews, audit existing tools for review analysis, and present a superior method for detecting review fraud. We hope that this research could be useful for e-commerce companies and other entities who have a stake in preserving opinion and review integrity online
Large Language Model Annotation Bias in Hate Speech Detection
No full textLarge language models (LLMs) are fast becoming ubiquitous and have shown impressive performance in various natural language processing (NLP) tasks. Annotating data for downstream applications is a resource-intensive task in NLP. Recently, the use of LLMs as a cost-effective data annotator for annotating data used to train other models or as an assistive tool has been explored. Yet, little is known regarding the societal implications of using LLMs for data annotation. In this work, focusing on hate speech detection, we investigate how using LLMs such as GPT-4 and Llama-3 for hate speech detection can lead to different performances for different text dialects and racial bias in online hate detection classifiers. We used LLMs to predict hate speech in seven hate speech datasets and trained classifiers on the LLM annotations of each dataset. Using tweets written in African-American English (AAE) and Standard American English (SAE), we show that classifiers trained on LLM annotations assign tweets written in AAE to negative classes (e.g., hate, offensive, abuse, racism, etc.) at a higher rate than tweets written in SAE and that the classifiers have a higher false positive rate towards AAE tweets. We explore the effect of incorporating dialect priming in the prompting techniques used in prediction, showing that introducing dialect increases the rate at which AAE tweets are assigned to negative classes
How Influencers and Multipliers Drive Polarization and Issue Alignment on Twitter/X
No full textWe investigate the polarization of the German Twittersphere by extracting the main issues discussed and the signaled opinions of users towards those issues based on (re)tweets concerning trending topics. The dataset covers daily trending topics from March 2021 to July 2023. At the opinion level, we show that the online public sphere is largely divided into two camps, one consisting mainly of left-leaning, and another of right-leaning accounts. Further we observe that political issues are strongly aligned, contrary to what one may expect from surveys. This alignment is driven by two cores of strongly active users: influencers, who generate ideologically charged content, and multipliers, who facilitate the spread of this content. The latter are specific to social media and play a crucial role as intermediaries on the platform by curating and amplifying very specific types of content that match their ideological position, resulting in the overall observation of a strongly polarized public sphere. These results contribute to a better understanding of the mechanisms that shape online public opinion, and have implications for the regulation of platforms
Perception of Visual Content: Differences Between Humans and Foundation Models
No full textHuman-annotated content is often used to train machine learning (ML) models. However, recently, language and multi-modal foundational models have been used to replace and scale-up human annotator's efforts. This study explores the similarity between human-generated and ML-generated annotations of images across diverse socio-economic contexts (RQ1) and their impact on ML model performance and bias (RQ2). We aim to understand differences in perception and identify potential biases in content interpretation. Our dataset comprises images of people from various geographical regions and income levels, covering various daily activities and home environments. ML captions and human labels show highest similarity at a low-level, i.e., types of words that appear and sentence structures, but all annotations are consistent in how they perceive images across regions. ML Captions resulted in best overall region classification performance, while ML Objects and ML Captions performed best overall for income regression. ML annotations worked best for action categories, while human input was more effective for non-action categories. These findings highlight the notion that both human and machine annotations are important, and that human-generated annotations are yet to be replaceable
Election Polls on Social Media: Prevalence, Biases, and Voter Fraud Beliefs
No full textSocial media platforms allow users to create polls to gather public opinion on diverse topics.
However, we know little about what such polls are used for and how reliable they are, especially in significant contexts like elections.
Focusing on the 2020 presidential elections in the U.S., this study shows that outcomes of election polls on Twitter deviate from election results despite their prevalence.
Leveraging demographic inference and statistical analysis, we find that Twitter polls are disproportionately authored by male Republicans and exhibit a large bias towards candidate Donald Trump in comparison to mainstream polls.
We investigate potential sources of biased outcomes from the point of view of inauthentic, automated, and counter-normative behavior. Using social media experiments and interviews with poll authors, we identify inconsistencies between public vote counts and those privately visible to poll authors, with the gap potentially attributable to purchased votes.
We find that election polls tend to be more biased, contain more questionable votes, and attract more bots before the election day than after. We highlight and compare key factors contributing to biased poll outcomes.
Finally, we identify instances of polls spreading voter fraud conspiracy theories and estimate that a couple of thousand such polls were posted in 2020.
The study discusses the implications of biased election polls in the context of transparency and accountability of social media platforms
Auditing Yelp’s Business Ranking and Review Recommendation Through the Lens of Fairness
No full textAuditing is critical to ensuring the fairness and reliability of decision-making systems. However, auditing a black-box system for bias can be challenging due to the lack of transparency in the model’s internal workings. In many web applications, such as Yelp, it is challenging, if not impossible, to manipulate their inputs systematically to identify bias in the output. Yelp connects users and businesses, where users identify new businesses and simultaneously express their experiences through reviews. Yelp recommendation software moderates user-provided content by categorizing it into recommended and not-recommended sections. The recommended reviews, among other attributes, are used by Yelp’s ranking algorithm to rank businesses in a neighborhood. Due to Yelp’s substantial popularity and its high impact on local businesses’ success, understanding the bias of its algorithms is crucial.
This data-driven study, for the first time, investigates the bias of Yelp’s business ranking and review recommendation system. We examine three hypotheses to assess if Yelp’s recommendation software shows bias against reviews of less established users with fewer friends and reviews and if Yelp’s business ranking algorithm shows bias against restaurants located in specific neighborhoods, particularly in hotspot regions, with specific demographic compositions. Our findings show that reviews of less-established users are disproportionately categorized as not-recommended. We also find a positive association between restaurants’ location in hotspot regions and their average exposure. Furthermore, we observed some cases of severe disparity bias in cities where the hotspots are in neighborhoods with less demographic diversity or higher affluence and education levels