Machine learning-based supervised classifiers are widely used for security tasks, and their improvement has been largely focused on algorithmic advancements. Data challenges that negatively impact the performance of these classifiers have received limited attention. We address the following research question: Can developments in Generative AI (GenAI) address data challenges and improve classifier performance? We propose augmenting training datasets with synthetic data generated using GenAI techniques to improve classifier generalization. We evaluate this approach across 7 diverse security tasks using 6 state-of-the-art GenAI methods and introduce a novel GenAI scheme called Nimai that enables highly controlled data synthesis. We find that GenAI techniques can significantly improve the performance of security classifiers, achieving improvements of up to 32.6% even in severely data-constrained settings (only ~180 training samples). Furthermore, we demonstrate that GenAI can facilitate rapid adaptation to concept drift post-deployment, requiring minimal labeling in the adjustment process. Despite successes, our study finds that some GenAI schemes struggle to initialize (train and produce data) on certain security tasks. We also identify characteristics of specific tasks, such as noisy labels, overlapping class distributions, and sparse feature vectors, which hinder performance boost using GenAI. We believe that our study will drive the development of future GenAI tools designed for security tasks.

Visual Privacy Protection (VPP) systems, such as blurring, pixelation, adversarial perturbations, and face replacement, are increasingly used to prevent identity disclosure in surveillance footage and publicly shared datasets. However, little is known about whether these protections apply equally across demographic groups, and uneven protection may create exploitable vulnerabilities. In this paper, we introduce FairDeFace, the first comprehensive framework for auditing the fairness, robustness, and visual quality of VPP systems under realistic adversarial threat models. Our framework evaluates how well different VPP techniques resist recognition attacks while preserving facial attributes and visual consistency across groups defined by gender and skin tone. It integrates benchmark datasets, face recognition models, attack pipelines, fairness measures, and novel visualization and statistical tools to trace where and why protections fail—for example, by identifying facial regions that are inconsistently altered. Across 500+ experiments involving seven state-of-the-art VPP methods, we uncover substantial disparities in protection strength and image quality. Certain groups, particularly Black males and females, receive consistently lower protection, framing fairness not only as an ethical concern but as a critical security and privacy risk. We release FairDeFace as an open and extensible framework to support the development and evaluation of equitable visual privacy systems.

Privacy policies are essential for communicating data practices to users, despite their dense, complex language, which often obstructs comprehension. Automated Policy Analyzer Tools (APATs) and recent advancements in Large Language Models (LLMs) have made strides in simplifying these documents through features such as question-answering and summarization. However, the robustness of these tools against adversarial manipulations remains less explored. This study examines how subtle, organization-side modifications to policy text can mislead APATs. We present APATRA, a multi-level framework for assessing the robustness of APATs against a range of policy-specific attacks designed to manipulate them into generating misleading outputs. We uncover vulnerabilities that can lead to substantial errors in outputs by applying character-, word-, phrase-, and sentence-level perturbations to policy contexts while preserving the original information. Our evaluation spans policy-specialized models (PolicyQA, PrivBERT), state-of-the-art LLMs (LLama-4-Maverick, GPT-3.5, GPT-4, GPT-5-mini, Claude-3.5, Claude-4.6-Sonnet), and third-party policy analyzer tools (AesirX). To validate the subtlety of our attacks, we conducted an IRB-approved user study with 100 general participants and 10 legal experts specializing in privacy law, who assessed adversarial excerpts for grammaticality, contradictions, and logical flow. Their evaluations demonstrate that our modifications maintain readability and raise no concerns. Overall, our results reveal critical security gaps in APATs and underscore the urgent need to enhance their robustness.

Auditing 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.

Social media platforms employ various content moderation techniques to remove harmful, offensive, and toxic content, with moderation levels varying across platforms and evolving over time. Parler, a fringe platform popular among conservative users, initially had minimal moderation, promoting itself as a space for open discussion. However, in 2021, it was removed from the Apple and Google App Stores and suspended from Amazon Web Services due to inadequate moderation of harmful content. After a month-long suspension, Parler returned with stricter guidelines, offering a unique opportunity to study the impact of platform-wide policy changes on user behavior and content outcomes. In this paper, we analyzed Parler data to assess the causal associations of these moderation changes on content toxicity and factuality. Using a longitudinal dataset of 17M posts from 432K users, who were active both before and after replatforming, we employed quasi-experimental analysis, controlling for confounding factors. We introduced a novel approach by using data from another social media platform, Twitter, to account for a critical confounding factor: offline events. This allowed us to isolate the effects of Parler's replatforming policies from external real-world influences. Our findings demonstrate that Parler's moderation changes are causally associated with a significant reduction in all forms of toxicity (p < 0.001). Additionally, we observed an increase in the factuality of the news sites shared and a reduction in the number of conspiracy/ pseudoscience sources.

Abortion remains one of the most controversial topics, especially after overturning Roe v. Wade ruling in the United States. Previous literature showed that the illegality of abortion could have serious consequences, as women might seek unsafe pregnancy terminations leading to increased maternal mortality rates and negative effects on their reproductive health. Therefore, the stances of the abortion-related Facebook posts were analyzed at the state level in the United States from May 4 until June 30, 2022, right after the Supreme Court’s decision was disclosed. In more detail, a pre-trained Transformer architecture-based model was fine-tuned on a manually labeled training set to obtain a stance detection model suitable for the collected dataset. Afterward, we employed appropriate statistical tests to examine the relationships between public opinion regarding abortion, abortion legality, political leaning, and factors measuring the overall population’s health, health knowledge, and vulnerability per state. We found that infant mortality rate, political affiliation, abortion rates, and abortion legality are associated with stances toward abortion at the state level in the US. While aligned with existing literature, these findings indicate how public opinion, laws, and women’s and infants’ health are related, as well as how these relationships can be demonstrated by using social media data.

Prior works have shown connections between online toxicity attacks, such as harassment, cyberbullying, and hate speech, and the subsequent increase in offline violence, as well as negative psychological effects on victims. These correlations are primarily identified through user studies conducted via virtual environments, simulations, and questionnaires. However, no work has investigated how, in practice and authentically, people react to online toxicity both emotionally, showing anger, anxiety, and sadness, and behaviorally in terms of engaging with and responding to toxicity instigators, considering conversations as a whole and the relation between emotions and behaviors. This data-driven study investigates the effect of toxicity on Twitter users' behaviors and emotions considering confounding factors, such as account identifiability, activity, and conversation's structure and topic. We collected about 80K Twitter conversations and identified those with and without toxic replies. Performing statistical tests along with propensity score matching, we investigated the causal association of receiving toxicity and users' responses. We found that authors of conversations with toxic replies are more likely to engage in conversations, reply in a toxic way, and unfollow toxicity instigators. In terms of users' emotional responses, we found that sadness and anger after the first toxic reply are more likely to increase as the amount of toxicity increases. These findings not only emphasize the negative emotional and behavioral effects of online toxicity on social media users but also, as demonstrated in this paper, can be utilized to build prediction models for users' reactions, which could then aid the implementation of proactive detection and intervention measures helping users in such situations.

The rapid growth in popularity and hype surrounding digital assets such as art, video, and music in the form of non-fungible tokens (NFTs) has made them a lucrative investment opportunity, with NFT-based sales surpassing $25B in 2021 alone. However, the volatility and general lack of technical understanding of the NFT ecosystem have led to the spread of various scams. The success of an NFT heavily depends on its online virality. As a result, creators use dedicated promotion services to drive engagement to their projects on social media websites, such as Twitter. However, these services are also utilized by scammers to promote fraudulent projects that attempt to steal users' cryptocurrency assets, thus posing a major threat to the ecosystem of NFT sales. In this paper, we conduct a longitudinal study of 439 promotion services (accounts) on Twitter that have collectively promoted 823 unique NFT projects through giveaway competitions over a period of two months. Our findings reveal that more than 36% of these projects were fraudulent, comprising of phishing, rug pull, and pre-mint scams. We also found that a majority of accounts engaging with these promotions (including those for fraudulent NFT projects) are bots that artificially inflate the popularity of the fraudulent NFT collections by increasing their likes, followers, and retweet counts. This manipulation results in significant engagement from real users, who then invest in these scams. We also identify several shortcomings in existing anti-scam measures, such as blocklists, browser protection tools, and domain hosting services, in detecting NFT-based scams. We utilize our findings to develop and open-source a machine learning classifier tool that was able to proactively detect 382 new fraudulent NFT projects on Twitter.

The advanced capabilities of Large Language Models (LLMs) have made them invaluable across various applications, from conversational agents and content creation to data analysis, research, and innovation. However, their effectiveness and accessibility also render them susceptible to abuse for generating malicious content, including phishing attacks. This study explores the potential of using four popular commercially available LLMs, i.e., ChatGPT (GPT 3.5 Turbo), GPT 4, Claude, and Bard, to generate functional phishing attacks using a series of malicious prompts. We discover that these LLMs can generate both phishing websites and emails that can convincingly imitate well-known brands and also deploy a range of evasive tactics that are used to elude detection mechanisms employed by anti-phishing systems. These attacks can be generated using unmodified or "vanilla" versions of these LLMs without requiring any prior adversarial exploits such as jailbreaking. We evaluate the performance of the LLMs towards generating these attacks and find that they can also be utilized to create malicious prompts that, in turn, can be fed back to the model to generate phishing scams - thus massively reducing the prompt-engineering effort required by attackers to scale these threats. As a countermeasure, we build a BERT-based automated detection tool that can be used for the early detection of malicious prompts to prevent LLMs from generating phishing content. Our model is transferable across all four commercial LLMs, attaining an average accuracy of 96% for phishing website prompts and 94% for phishing email prompts. We also disclose the vulnerabilities to the concerned LLMs, with Google acknowledging it as a severe issue. Our detection model is available for use at Hugging Face, as well as a ChatGPT Actions plugin.

Social media platforms have been establishing content moderation guidelines and employing various moderation policies to counter hate speech and misinformation. The goal of this paper is to study these community guidelines and moderation practices, as well as the relevant research publications, to identify the research gaps, differences in moderation techniques, and challenges that should be tackled by the social media platforms and the research community. To this end, we study and analyze fourteen most popular social media content moderation guidelines and practices, and consolidate them. We then introduce three taxonomies drawn from this analysis as well as covering over two hundred interdisciplinary research papers about moderation strategies. We identify the differences between the content moderation employed in mainstream and fringe social media platforms. Finally, we have in-depth applied discussions on both research and practical challenges and solutions.

Prior work has extensively studied misinformation related to news, politics, and health, however, misinformation can also be about technological topics. While less controversial, such misinformation can severely impact companies’ reputations and revenues, and users’ online experiences. Recently, social media has also been increasingly used as a novel source of knowledgebase for extracting timely and relevant security threats, which are fed to the threat intelligence systems for better performance. However, with possible campaigns spreading false security threats, these systems can become vulnerable to poisoning attacks. In this work, we proposed novel approaches for detecting misinformation about cybersecurity and privacy threats on social media, focusing on two topics with different types of misinformation: phishing websites and Zoom’s security & privacy threats. We developed a framework for detecting inaccurate phishing claims on Twitter. Using this framework, we could label about 9% of URLs and 22% of phishing reports as misinformation. We also proposed another framework for detecting misinformation related to Zoom’s security and privacy threats on multiple platforms. Our classifiers showed great performance with more than 98% accuracy. Employing these classifiers on the posts from Facebook, Instagram, Reddit, and Twitter, we found respectively that about 18%, 3%, 4%, and 3% of posts were misinformation. In addition, we studied the characteristics of misinformation posts, their authors, and their timelines, which helped us identify campaigns.

Free Website Building services (FWBs) provide individuals with a cost-effective and convenient way to create a website without requiring advanced technical knowledge or coding skills. However, malicious actors often abuse these services to host phishing websites. In this work, we propose FreePhish, a scalable framework to continuously identify phishing websites that are created using FWBs. Using FreePhish, we were able to detect and characterize more than 31.4K phishing URLs that were created using 17 unique free website builder services and shared on Twitter and Facebook over a period of six months. We find that FWBs provide attackers with several features that make it easier to create and maintain phishing websites at scale while simultaneously evading anti-phishing countermeasures. Our study indicates that anti-phishing blocklists and browser protection tools have significantly lower coverage and high detection time against FWB phishing attacks when compared to regular (self-hosted) phishing websites. While our prompt disclosure of these attacks helped some FWBs to remove these attacks, we found several others who were slow at removal or did not remove them outright, with the same also being true for Twitter and Facebook. Finally, we also provide FreePhish as a free Chromium web extension that can be utilized to prevent end-users from accessing potential FWB-based phishing attacks.

Several face de-identification methods have been proposed to preserve users’ privacy by obscuring their faces. These methods, however, can degrade the quality of photos, and they usually do not preserve the utility of faces, i.e., their age, gender, pose, and facial expression. Recently, advanced generative adversarial network models, such as StyleGAN [ 33], have been proposed, which generate realistic, high-quality imaginary faces. In this paper, we investigate the use of StyleGAN in generating de-identified faces through style mixing, where the styles or features of the target face and an auxiliary face get mixed to generate a de-identified face that carries the utilities of the target face. We examined this de-identification method for preserving utility and privacy by implementing several face detection, verification, and identification attacks and conducting a user study. The results from our extensive experiments, human evaluation, and comparison with two state-of-the-art face de-identification methods, i.e., CIAGAN and DeepPrivacy, show that StyleGAN performs on par or better than these methods, preserving users’ privacy and images’ utility. In particular, the results of the machine learning-based experiments show that StyleGAN0-4 preserves utility better than CIAGAN and DeepPrivacy while preserving privacy at the same level. StyleGAN 0-3 preserves utility at the same level while providing more privacy. In this paper, for the first time, we also performed a carefully designed user study to examine both privacy and utility-preserving properties of StyleGAN 0-3, 0-4, and 0-5, as well as CIAGAN and DeepPrivacy from the human observers’ perspectives. Our statistical tests showed that participants tend to verify and identify StyleGAN 0-5 images easier than DeepPrivacy images. All the methods but StyleGAN 0-5 had significantly lower identification rates than CIAGAN. Regarding utility, as expected, StyleGAN 0-5 performed significantly better in preserving some attributes. Among all methods, on average, participants believe gender has been preserved the most while naturalness has been preserved the least.

Side-channel attacks allow an adversary to uncover secret program data by observing the behavior of a program with respect to a resource, such as execution time, consumed memory or response size. Side-channel vulnerabilities are difficult to reason about as they involve analyzing the correlations between resource usage over multiple program paths. We present DifFuzz, a fuzzing-based approach for detecting side-channel vulnerabilities related to time and space. DifFuzz automatically detects these vulnerabilities by analyzing two versions of the program and using resource-guided heuristics to find inputs that maximize the difference in resource consumption between secret-dependent paths. The methodology of DifFuzz is general and can be applied to programs written in any language. For this paper, we present an implementation that targets analysis of Java programs, and uses and extends the Kelinci and AFL fuzzers. We evaluate DifFuzz on a large number of Java programs and demonstrate that it can reveal unknown side-channel vulnerabilities in popular applications. We also show that DifFuzz compares favorably against Blazer and Themis, two state-of-the-art analysis tools for finding side-channels in Java programs.

While online review services provide a two-way conversation between brands and consumers, malicious actors, including misbehaving businesses, have an equal opportunity to distort the reviews for their own gains. We propose OneReview, a method for locating fraudulent reviews, correlating data from multiple crowd-sourced review sites. Our approach utilizes Change Point Analysis to locate points at which a business' reputation shifts. Inconsistent trends in reviews of the same businesses across multiple websites are used to identify suspicious reviews. We then extract an extensive set of textual and contextual features from these suspicious reviews and employ supervised machine learning to detect fraudulent reviews.

Cybercriminals have found in online social networks a propitious medium to spread spam and malicious content. Existing techniques for detecting spam include predicting the trustworthiness of accounts and analyzing the content of these messages. However, advanced attackers can still successfully evade these defenses.

Online social network providers have become treasure troves of information for marketers and researchers. To profit from their data while honoring the privacy of their customers, social networking services share `anonymized' social network datasets, where, for example, identities of users are removed from the social network graph. However, by using external information such as a reference social graph (from the same network or another network with similar users), researchers have shown how such datasets can be de-anonymized. These approaches use `network alignment' techniques to map nodes from the reference graph into the anonymized graph and are often sensitive to larger network sizes, the number of seeds, and noise --- which may be added to preserve privacy. We propose a divide-and-conquer approach to strengthen the power of such algorithms. Our approach partitions the networks into `communities' and performs a two-stage mapping: first at the community level, and then for the entire network. Through extensive simulation on real-world social network datasets, we show how such community-aware network alignment improves de-anonymization performance under high levels of noise, large network sizes, and a low number of seeds. Even when nodes cannot be explicitly mapped, the community structure can be mapped between both networks, thus reducing the anonymity of users. For example, for our (real-world) Twitter dataset with 90,000 nodes, 20% noise, and 16 seeds, the state-of-the-art technique reduces anonymity by 0 bits, whereas our approach reduces anonymity by 9.71 bits (with 40% of nodes mapped).

Online social networks (OSNs) such as Facebook and Google+ have transformed the way our society communicates. However, this success has come at the cost of user privacy; in today's OSNs, users are not in control of their own data, and depend on OSN operators to enforce access control policies. A multitude of privacy breaches has spurred research into privacy-preserving alternatives for social networking, exploring a number of techniques for storing, disseminating, and controlling access to data in a decentralized fashion. In this paper, we argue that a combination of techniques is necessary to efficiently support the complex functionality requirements of OSNs.