Large Language Models (LLMs) have demonstrated strong capabilities as autonomous agents through tool use, planning, and decisionmaking abilities, leading to their widespread adoption across diverse tasks.As task complexity grows, multi-agent LLM systems are increasingly used to solve problems collaboratively.However, safety and security of these systems remains largely under-explored.Existing benchmarks and datasets predominantly focus on single-agent settings, failing to capture the unique vulnerabilities of multi-agent dynamics and co-ordination.To address this gap, we introduce Threats and Attacks in Multi-Agent Systems (TAMAS), a benchmark designed to evaluate the robustness and safety of multi-agent LLM systems.TAMAS includes five distinct scenarios comprising 300 adversarial instances across six attack types and 211 tools, along with 100 harmless tasks.We assess system performance across ten backbone LLMs and three agent interaction configurations from Autogen and CrewAI frameworks, highlighting critical challenges and failure modes in current multi-agent deployments.Furthermore, we introduce Effective Robustness Score (ERS) to assess the tradeoff between safety and task effectiveness of these frameworks.Our findings show that multi-agent systems are highly vulnerable to adversarial attacks, underscoring the urgent need for stronger defenses.TAMAS provides a foundation for systematically studying and improving the safety of multi-agent LLM systems.Code and dataset is available at https://github.com/microsoft/TAMAS.

Graph Neural Networks (GNNs) have emerged as powerful tools for learning representations of graph-structured data, demonstrating remarkable performance across various tasks. Recognizing their importance, there has been extensive research focused on explaining GNN predictions, aiming to enhance their interpretability and trustworthiness. However, GNNs and their explainers face a notable challenge: graphs are primarily designed to model pair-wise relationships between nodes, which can make it tough to capture higher-order, multi-node interactions. This characteristic can pose difficulties for existing explainers in fully representing multi-node relationships. To address this gap, we present Framework For Higher-Order Representations In Graph Explanations (FORGE), a framework that enables graph explainers to capture such interactions by incorporating higher-order structures, resulting in more accurate and faithful explanations. Extensive evaluation shows that on average real-world datasets from the GraphXAI benchmark and synthetic datasets across various graph explainers, FORGE improves average explanation accuracy by 1.9x and 2.25x, respectively. We perform ablation studies to confirm the importance of higher-order relations in improving explanations, while our scalability analysis demonstrates FORGE's efficacy on large graphs.

Making analogies is fundamental to cognition. Proportional analogies, which consist of four terms, are often used to assess linguistic and cognitive abilities. For instance, completing analogies like “Oxygen is to Gas as < blank > is to < blank >" requires identifying the semantic relationship (e.g., “type of”) between the first pair of terms (“Oxygen” and “Gas”) and finding a second pair that shares the same relationship (e.g., “Aluminum” and “Metal”). In this work, we introduce a 15K Multiple-Choice Question Answering (MCQA) dataset for proportional analogy completion and evaluate the performance of contemporary Large Language Models (LLMs) in various knowledge-enhanced prompt settings. Specifically, we augment prompts with three types of knowledge: exemplar, structured, and targeted. Our results show that despite extensive training data, solving proportional analogies remains challenging for current LLMs, with the best model achieving an accuracy of 55%. Notably, we find that providing targeted knowledge can better assist models in completing proportional analogies compared to providing exemplars or collections of structured knowledge. Our code and data are available at: https://github.com/Thiliniiw/KnowledgePrompts/

This paper investigates advertising practices in print newspapers across India using a novel data-driven approach.We develop a pipeline employing image processing and OCR techniques to extract articles and advertisements from digital versions of print newspapers with high accuracy.Applying this methodology to five popular newspapers that span multiple regions and three languages-English, Hindi, and Telugu-we assembled a dataset of more than 12,000 editions containing several hundred thousand advertisements.Collectively, these newspapers reach a readership of over 100 million people.Using this extensive dataset, we conduct a comprehensive analysis to answer key questions about print advertising: who advertises, what they advertise, when they advertise, where they place their ads, and how they advertise.Our findings reveal significant patterns, including the consistent level of print advertising over the past six years despite declining print circulation, the overrepresentation of company ads on prominent pages, and the disproportionate revenue contributed by government ads.Furthermore, we examine whether advertising in a newspaper influences the coverage an advertiser receives.Through regression analyses on coverage volume and sentiment, we find strong evidence supporting this hypothesis for corporate advertisers.The results indicate a clear trend where increased advertising correlates with more favorable and extensive media coverage, a relationship that remains robust over time and across different levels of advertiser popularity.

Function calling is a complex task with widespread applications in domains such as information retrieval, software engineering and automation. For example, a query to book the shortest flight from New York to London on January 15 requires identifying the correct parameters to generate accurate function calls. Large Language Models (LLMs) can automate this process but are computationally expensive and impractical in resource-constrained settings. In contrast, Small Language Models (SLMs) can operate efficiently, offering faster response times, and lower computational demands, making them potential candidates for function calling on edge devices. In this exploratory empirical study, we evaluate the efficacy of SLMs in generating function calls across diverse domains using zero-shot, few-shot, and fine-tuning approaches, both with and without prompt injection, while also providing the finetuned models to facilitate future applications. Furthermore, we analyze the model responses across a range of metrics, capturing various aspects of function call generation. Additionally, we perform experiments on an edge device to evaluate their performance in terms of latency and memory usage, providing useful insights into their practical applicability. Our findings show that while SLMs improve from zero-shot to few-shot and perform best with fine-tuning, they struggle significantly with adhering to the given output format. Prompt injection experiments further indicate that the models are generally robust and exhibit only a slight decline in performance. While SLMs demonstrate potential for the function call generation task, our results also highlight areas that need further refinement for real-time functioning.

Arvindh Arun, Sumit Kumar, Mojtaba Nayyeri, Bo Xiong, Ponnurangam Kumaraguru, Antonio Vergari, Steffen Staab. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Seyedali Mohammadi, Bhaskara Hanuma Vedula, Hemank Lamba, Edward Raff, Ponnurangam Kumaraguru, Francis Ferraro, Manas Gaur. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

As Language Model (LM) capabilities advance, evaluating and supervising them at scale is getting harder for humans. There is hope that other language models can automate both these tasks, which we refer to as AI Oversight. We study how model similarity affects both aspects of AI oversight by proposing Chance Adjusted Probabilistic Agreement (CAPA)–a metric for LM similarity based on overlap in model mistakes. Using CAPA, we first show that LLM-as-a-judge scores favor models similar to the judge, generalizing recent self-preference results. Then, we study training on LM annotations, and find complementary knowledge between the weak supervisor and strong student model plays a crucial role in gains from weak-to-strong generalization. As model capabilities increase, it becomes harder to find their mistakes, and we might defer more to AI oversight. However, we observe a concerning trend–model mistakes are becoming more similar with increasing capabilities, pointing to risks from correlated failures. Our work underscores the importance of reporting and correcting for model similarity, especially in the emerging paradigm of AI oversight.

Graph Neural Networks (GNNs) are increasingly being used for a variety of ML applications on graph data. Because graph data does not follow the independently and identically distributed i.i.d. assumption, adversarial manipulations or incorrect data can propagate to other data points through message passing, which deteriorates the model’s performance. To allow model developers to remove the adverse effects of manipulated entities from a trained GNN, we study the recently formulated problem of Corrective Unlearning. We find that current graph unlearning methods fail to unlearn the effect of manipulations even when the whole manipulated set is known. We introduce a new graph unlearning method,Cognac, which can unlearn the effect of the manipulation set even when only $5$% of it is identified. It recovers most of the performance of a strong oracle with fully corrected training data, even beating retraining from scratch without the deletion set, and is $8$x more efficient while also scaling to large datasets. We hope our work assists GNN developers in mitigating harmful effects caused by issues in real-world data, post-training.

Current computational approaches for analysing or generating code-mixed sentences do not explicitly model “naturalness” or “acceptability” of code-mixed sentences, but rely on training corpora to reflect distribution of acceptable code-mixed sentences. Modelling human judgement for the acceptability of code-mixed text can help in distinguishing natural code-mixed text and enable quality-controlled generation of code-mixed text. To this end, we construct Cline—a dataset containing human acceptability judgements for English-Hindi (en-hi) code-mixed text. Cline is the largest of its kind with 16,642 sentences, consisting of samples sourced from two sources: synthetically generated code-mixed text and samples collected from online social media. Our analysis establishes that popular code-mixing metrics such as CMI, Number of Switch Points, Burstines, which are used to filter/curate/compare code-mixed corpora have low correlation with human acceptability judgements, underlining the necessity of our dataset. Experiments using Cline demonstrate that simple Multilayer Perceptron (MLP) models when trained solely using code-mixing metrics as features are outperformed by fine-tuned pre-trained Multilingual Large Language Models (MLLMs). Specifically, among Encoder models XLM-Roberta and Bernice outperform IndicBERT across different configurations. Among Encoder-Decoder models, mBART performs better than mT5, however, Encoder-Decoder models are not able to outperform Encoder-only models. Decoder-only models perform the best when compared with all other MLLMS, with Llama 3.2 - 3B models outperforming similarly sized Qwen, Phi models. Comparison with zero and fewshot capabilitites of ChatGPT show that MLLMs fine-tuned on larger data outperform ChatGPT, providing scope for improvement in code-mixed tasks. Zero-shot transfer from English–Hindi to English-Telugu acceptability judgments using our model checkpoints proves superior to random baselines, enabling application to other code-mixed language pairs and providing further avenues of research. We publicly release our human-annotated dataset, trained checkpoints, code-mix corpus, and code for data generation and model training.

Despite recent advancements showcasing the impressive capabilities of Large Language Models (LLMs) in conversational systems, we show that even state-of-the-art LLMs are morally inconsistent in their generations, questioning their reliability (and trustworthiness in general). Prior works in LLM evaluation focus on developing ground-truth data to measure accuracy on specific tasks. However, for moral scenarios that often lack universally agreed-upon answers, consistency in model responses becomes crucial for their reliability. To address this issue, we propose an information-theoretic measure called Semantic Graph Entropy (SaGE), grounded in the concept of “Rules of Thumb” (RoTs) to measure a model’s moral consistency. RoTs are abstract principles learned by a model and can help explain their decision-making strategies effectively. To this extent, we construct the Moral Consistency Corpus (MCC), containing 50K moral questions, responses to them by LLMs, and the RoTs that these models followed. Furthermore, to illustrate the generalizability of SaGE, we use it to investigate LLM consistency on two popular datasets – TruthfulQA and HellaSwag. Our results reveal that task accuracy and consistency are independent problems, and there is a dire need to investigate these issues further.

Coreference resolution involves the task of identifying text spans within a discourse that pertain to the same real-world entity. While this task has been extensively explored in the English language, there has been a notable scarcity of publicly accessible resources and models for coreference resolution in South Asian languages. We introduce a Translated dataset for Multilingual Coreference Resolution (TransMuCoRes) in 31 South Asian languages using off-the-shelf tools for translation and word-alignment. Nearly all of the predicted translations successfully pass a sanity check, and 75% of English references align with their predicted translations. Using multilingual encoders, two off-the-shelf coreference resolution models were trained on a concatenation of TransMuCoRes and a Hindi coreference resolution dataset with manual annotations. The best performing model achieved a score of 64 and 68 for LEA F1 and CoNLL F1, respectively, on our test-split of Hindi golden set. This study is the first to evaluate an end-to-end coreference resolution model on a Hindi golden set. Furthermore, this work underscores the limitations of current coreference evaluation metrics when applied to datasets with split antecedents, advocating for the development of more suitable evaluation metrics.

The White House Executive Order on Artificial Intelligence highlights the risks of large language models (LLMs) empowering malicious actors in developing biological, cyber, and chemical weapons. To measure these risks, government institutions and major AI labs are developing evaluations for hazardous capabilities in LLMs. However, current evaluations are private and restricted to a narrow range of malicious use scenarios, which limits further research into reducing malicious use. To fill these gaps, we release the Weapons of Mass Destruction Proxy (WMDP) benchmark, a dataset of 3,668 multiple-choice questions that serve as a proxy measurement of hazardous knowledge in biosecurity, cybersecurity, and chemical security. To guide progress on unlearning, we develop RMU, a state-of-the-art unlearning method based on controlling model representations. RMU reduces model performance on WMDP while maintaining general capabilities in areas such as biology and computer science, suggesting that unlearning may be a concrete path towards reducing malicious use from LLMs. We release our benchmark and code publicly at https://wmdp.ai.

Language models often exhibit undesirable behavior, e.g., generating toxic or gender-biased text. In the case of neural language models, an encoding of the undesirable behavior is often present in the model’s representations. Thus, one natural (and common) approach to prevent the model from exhibiting undesirable behavior is to steer the model’s representations in a manner that reduces the probability of it generating undesirable text. This paper investigates the formal and empirical properties of steering functions, i.e., transformation of the neural language model’s representations that alter its behavior. First, we derive two optimal, in the least-squares sense, affine steering functions under different constraints. Our theory provides justification for existing approaches and offers a novel, improved steering approach. Second, we offer a series of experiments that demonstrate the empirical effectiveness of the methods in mitigating bias and reducing toxic generation.

Social media platforms, particularly Twitter, have emerged as vital media for organizing online protests worldwide. During protests, users on social media share different narratives, often coordinated to share collective opinions and obtain widespread reach. In this paper, we focus on the communities formed during a protest and the collective narratives they share, using the protest on the enactment of the Citizenship Amendment Act (#CAA) by the Indian Government as a case study. Since #CAA protest led to divergent discourse in the country, we first classify the users into opposing stances, i.e., protesters (who opposed the Act) and counter-protesters (who supported it) in an unsupervised manner. Next, we identify the coordinated communities in the opposing stances and examine the collective narratives shared by coordinated communities of opposing stances. We use content-based metrics to identify user coordination, including hashtags, mentions, and retweets. Our results suggest mention as the strongest metric for coordination across the opposing stances. Next, we decipher the collective narratives in the opposing stances using an unsupervised narrative detection framework and found call-to-action, on-ground activity, grievances sharing, questioning, and skepticism narratives in the protest tweets. We analyze the strength of the different coordinated communities using network measures, and perform inauthentic activity analysis on the most coordinated communities on both sides. Our findings also suggest that coordinated communities, which were highly inauthentic, showed the highest clustering coefficient towards a greater extent of coordination.

The default approach to deal with the enormous size and limited accessibility of many Web and social media networks is to sample one or more subnetworks from a conceptually unbounded unknown network. Clearly, the extracted subnetworks will crucially depend on the sampling scheme. Motivated by studies of homophily and opinion formation, we propose a variant of snowball sampling designed to prioritize the inclusion of entire cohesive communities rather than any kind of representativeness, breadth, or depth of coverage. The method is illustrated on a concrete example, and experiments on synthetic networks suggest that it behaves as desired.

The pursuit of habit building is challenging, and most people struggle with it. Research on successful habit formation is mainly based on small human trials focusing on the same habit for all the participants as conducting long-term heterogonous habit studies can be logistically expensive. With the advent of self-help, there has been an increase in online communities and applications that are centered around habit building and logging. Habit building applications can provide large-scale data on real-world habit building attempts and unveil the commonalities among successful ones. We collect public data on stickk.com, which allows users to track progress on habit building attempts called commitments. A commitment can have an external referee, regular check-ins about the progress, and a monetary stake in case of failure. Our data consists of 742,923 users and 397,456 commitments. In addition to the dataset, rooted in theories like Fresh Start Effect, Accountablity, and Loss Aversion, we ask questions about how commitment properties like start date, external accountability, monitory stake, and pursuing multiple habits together affects the odds of success. We found that people tend to start habits on temporal landmarks, but that does not affect the probability of their success. Practices like accountability and stakes are not often used but are strong determents of success. Commitments of 6 to 8 weeks in length, weekly reporting with an external referee, and a monetary amount at stake tend to be most successful. Finally, around 40% of all commitments are attempted simultaneously with other goals. Simultaneous attempts of pursuing commitments may fail early, but if pursued through the initial phase, they are statistically more successful than building one habit at a time.

Large Language Models (LLMs) have emerged as powerful tools to perform various tasks in the legal domain, ranging from generating summaries to predicting judgments. Despite their immense potential, these models have been proven to learn and exhibit societal biases and make unfair predictions. Hence, it is essential to evaluate these models prior to deployment. In this study, we explore the ability of LLMs to perform Binary Statutory Reasoning in the Indian legal landscape across various societal disparities. We present a novel metric, β-weighted Legal Safety Score (LSSβ), to evaluate the legal usability of the LLMs. Additionally, we propose a finetuning pipeline, utilising specialised legal datasets, as a potential method to reduce bias. Our proposed pipeline effectively reduces bias in the model, as indicated by improved LSSβ. This highlights the potential of our approach to enhance fairness in LLMs, making them more reliable for legal tasks in socially diverse contexts.

In this paper, we tackle the complex task of analyzing televised debates, with a focus on a prime time news debate show from India. Previous methods, which often relied solely on text, fall short in capturing the multimodal essence of these debates [27]. To address this gap, we introduce a comprehensive automated toolkit that employs advanced computer vision and speech-to-text techniques for large-scale multimedia analysis. Utilizing state-of-the-art computer vision algorithms and speech-to-text methods, we transcribe, diarize, and analyze thousands of YouTube videos of a prime-time television debate show in India. These debates are a central part of Indian media but have been criticized for compromised journalistic integrity and excessive dramatization [18]. Our toolkit provides concrete metrics to assess bias and incivility, capturing a comprehensive multimedia perspective that includes text, audio utterances, and video frames. Our findings reveal significant biases in topic selection and panelist representation, along with alarming levels of incivility. This work offers a scalable, automated approach for future research in multimedia analysis, with profound implications for the quality of public discourse and democratic debate. To catalyze further research in this area, we also release the code, dataset collected and supplemental pdf.1

Continual learning has primarily focused on the issue of catastrophic forgetting and the associated stability-plasticity tradeoffs. However, little attention has been paid to the efficacy of continually learned representations, as representations are learned alongside classifiers throughout the learning process. Our primary contribution is empirically demonstrating that existing online continually trained deep networks produce inferior representations compared to a simple pre-defined random transforms. Our approach embeds raw pixels using a fixed random transform, approximating an RBF-Kernel initialized before any data is seen. We then train a simple linear classifier on top without storing any exemplars, processing one sample at a time in an online continual learning setting. This method, called RanDumb, significantly outperforms state-of-the-art continually learned representations across all standard online continual learning benchmarks. Our study reveals the significant limitations of representation learning, particularly in low-exemplar and online continual learning scenarios. Extending our investigation to popular exemplar-free scenarios with pretrained models, we find that training only a linear classifier on top of pretrained representations surpasses most continual fine-tuning and prompt-tuning strategies. Overall, our investigation challenges the prevailing assumptions about effective representation learning in the online continual learning.

The COVID-19 pandemic that broke out globally in December 2019 put us all in an unprecedented situation. Social media became a vital source of support and information during the pandemic, as physical interactions were limited by people staying at home. This paper investigates support dynamics and user commitment in an online COVID-19 community of Reddit. We define various support classes and observe them along with user behavior and temporal phases for a coherent in the community. We perform survival analysis using Cox Regression to identify factors influencing a user's commitment to the community. People seeking more emotional and informational support while they are COVID-positive stay longer in the community. Surprisingly, people who give more support in their early phases are less likely to stay. Additionally, contrary to common belief, our findings show that receiving emotional and informational support has little effect on users' longevity in the community. Our results lead to a better understanding of user dynamics related to community support and can directly impact moderators and platform owners in designing community guidelines and incentive structures.

Unsupervised Representation Learning on graphs is gaining traction due to the increasing abundance of unlabelled network data and the compactness, richness, and usefulness of the representations generated. In this context, the need to consider fairness and bias constraints while generating the representations has been well-motivated and studied to some extent in prior works. One major limitation of most of the prior works in this setting is that they do not aim to address the bias generated due to connectivity patterns in the graphs, such as varied node centrality, which leads to a disproportionate performance across nodes. In our work, we aim to address this issue of mitigating bias due to inherent graph structure in an unsupervised setting. To this end, we propose CAFIN, a centrality-aware fairness-inducing framework that leverages the structural information of graphs to tune the representations generated by existing frameworks. We deploy it on GraphSAGE (a popular framework in this domain) and showcase its efficacy on two downstream tasks – Node Classification and Link Prediction. Empirically, CAFIN consistently reduces the performance disparity across popular datasets (varying from 18 to 80% reduction in performance disparity) from various domains while incurring only a minimal cost of fairness.

Deaths due to drug overdose in the US have doubled in the last decade. Drug-related content on social media has also exploded in the same time frame. The pseudo-anonymous nature of social media platforms enables users to discourse about taboo and sometimes illegal topics like drug consumption. User-generated content (UGC) about drugs on social media can be used as an online proxy to detect offline drug consumption. UGC also gets exposed to the praise and criticism of the community. Law of effect proposes that positive reinforcement on an experience can incentivize the users to engage in the experience repeatedly. Therefore, we hypothesize that positive community feedback on a user's online drug consumption disclosure will increase the probability of the user doing an online drug consumption disclosure post again. To this end, we collect data from 10 drug-related subreddits. First, we build a deep learning model to classify UGC as indicative of drug consumption offline or not, and analyze the extent of such activities. Further, we use matching-based causal inference techniques to unravel community feedback's effect on users' future drug consumption behavior. We discover that 84% of posts and 55% comments on drug-related subreddits indicate real-life drug consumption. Users who get positive feedback generate up to two times more drugs consumption content in the future. Finally, we conducted an anonymous user study on drug-related subreddits to compare members' opinions with our experimental findings and show that user tends to underestimate the effect community peers can have on their decision to interact with drugs.

Complex legal language, filled with jargon, nuanced language semantics, and a high level of domain specificity, poses a significant challenge for automation in handling various legal tasks. In the realm of legal document composition, a pivotal component revolves around accurately referencing case laws and other sources to substantiate assertions and arguments. Understanding the legal domain and identifying appropriate citation context or cite-worthy sentences automatically is challenging. Our research is centered on the issue of citation-worthiness identification of a given sentence. This serves as the initial phase in contemporary citation recommendation systems, aimed at alleviating the effort involved in extracting a suitable array of citation contexts. To address this, we first introduce a labeled dataset comprising 178 million sentences, specifically tailored for detecting citation-worthy content within the legal domain. This dataset is curated from the Caselaw Access Project (CAP) (https://case.law/). We proceeded to assess the performance of a range of deep learning models on this novel dataset. Among the models examined, the domain-specific pre-trained model consistently demonstrated superior performance, achieving an 88% F1-score in the task of detecting citation-worthy material. To enhance our insights, we employed inputXGradient explainable AI techniques to dissect the predictions, thereby identifying the tokens that contribute to specific citation classes.

Suspect Identification is one of the most pivotal aspects of a forensic and criminal investigation. A significant amount of time and skill is devoted to creating sketches for it and requires a fair amount of recollections from the witness to provide a useful sketch. We devise a method that aims to automate the process of suspect identification and model this problem by iteratively retrieving images from feedback provided by the user. Compared to standard image retrieval tasks, interactive facial image retrieval is specifically more challenging due to the high subjectivity involved in describing a person’s facial attributes and appropriately evolving with the preferences put forward by the user. Our method uses a relatively simpler form of supervision by utilizing the user’s feedback to label images as either similar or dissimilar to their mental image of the suspect based on which we propose a loss function using the contrastive learning paradigm that is optimized in an online fashion. We validate the efficacy of our proposed approach using a carefully designed testbed to simulate user feedback and a large-scale user study. We empirically show that our method iteratively improves personalization, leading to faster convergence and enhanced recommendation relevance, thereby, improving user satisfaction. Our proposed framework is being designed for real-time use in the metropolitan crime investigation department, and thus is also equipped with a user-friendly web interface with a real-time experience for suspect retrieval.

Targeted image retrieval has long been a challenging problem since each person has a different perception of different features leading to inconsistency among users in describing the details of a particular image. Due to this, each user needs a system personalized according to the way they have structured the image in their mind. One important application of this task is suspect identification in forensic investigations where a witness needs to identify the suspect from an existing criminal database. Existing methods require the attributes for each image or suffer from poor latency during training and inference. We propose a new approach to tackle this problem through explicit relevance feedback by introducing a novel loss function and a corresponding scoring function. For this, we leverage contrastive learning on the user feedback to generate the next set of suggested images while improving the level of personalization with each user feedback iteration.

Protests and mass mobilization are scarce; however, they may lead to dramatic outcomes when they occur. Social media such as Twitter has become a center point for the organization and development of online protests worldwide. It becomes crucial to decipher various narratives shared during an online protest to understand people's perceptions. In this work, we propose an unsupervised clustering-based framework to understand the narratives present in a given online protest. Through a comparative analysis of tweet clusters in 3 protests around government policy bills, we contribute novel insights about narratives shared during an online protest. Across case studies of government policy-induced online protests in India and the United Kingdom, we found familiar mass mo-bilization narratives across protests. We found reports of on-ground activities and call-to-action for people's participation narrative clusters in all three protests under study. We also found protest-centric narratives in different protests, such as skepticism around the topic. The results from our analysis can be used to understand and compare people's perceptions of future mass mobilizations.

Non-financial incentives in the form of awards often act as a driver of positive reinforcement and elevation of social status in the offline world. The elevated social status results in people becoming more active, aligning to a change in the communities' expectations. However, the impact in terms of longevity of social influence and community acceptance of leaders of these incentives in the form of awards are not well-understood in the online world. Our work aims to shed light on the impact of these awards on the awardee and the community. We focus on three large subreddits with a snapshot of 219K posts and 5.8 million comments contributed by 88K Reddit users who received 14,146 awards. Our work establishes that the behaviour of awardees change statistically significantly for a short time after getting an award; however, the change is ephemeral since the awardees return to their pre-award behaviour within days. Additionally, via a user survey, we identified a long-lasting impact of awards-we found that the community's stance softened towards awardees.

With Twitter growing as a preferred channel for outreach among major politicians, there have been focused efforts on online communication, even in election campaigns in primarily rural regions. In this paper, we examine the relationship between politicians’ use of social media and the level of urbanization and literacy by compiling a comprehensive list of Twitter handles of political party functionaries and election candidates in the run-up to the 2022 State Assembly elections in Uttar Pradesh, India. We find statistically significant relationships between political Twitter presence and levels of urbanization and with levels of literacy. We also find a strong correlation between vote share and Twitter presence in the winning party, a relationship that is even stronger in urban districts. This provides empirical evidence that social media is already a central part of electoral outreach processes in the Global South, but that this is still selectively more relevant to voters in, and politicians standing for elections from urban and higher-educated regions.

Polysemy is the phenomenon where a single word form possesses two or more related senses.It is an extremely ubiquitous part of natural language and analyzing it has sparked rich discussions in the linguistics, psychology and philosophy communities alike.With scarce attention paid to polysemy in computational linguistics, and even scarcer attention toward quantifying polysemy, in this paper, we propose a novel, unsupervised framework to compute and estimate polysemy scores for words in multiple languages.We infuse our proposed quantification with syntactic knowledge in the form of dependency structures.This informs the final polysemy scores of the lexicon motivated by recent linguistic findings that suggest there is an implicit relation between syntax and ambiguity/polysemy.We adopt a graph based approach by computing the discrete Ollivier Ricci curvature on a graph of the contextual nearest neighbors.We test our framework on curated datasets controlling for different sense distributions of words in 3 typologically diverse languages -English, French and Spanish.The effectiveness of our framework is demonstrated by significant correlations of our quantification with expert human annotated language resources like WordNet.We observe a 0.3 point increase in the correlation coefficient as compared to previous quantification studies in English.Our research leverages contextual language models and syntactic structures to empirically support the widely held theoretical linguistic notion that syntax is intricately linked to ambiguity/polysemy.

In the last two decades, Information and Communication Technologies (ICTs) have played a pivotal role in empowering rural populations in India by making knowledge more accessible. Digital Green is one such ICT that employs a participatory approach with smallholder farmers to produce instructional agricultural videos that encompass content specific to them. With the help of human mediators, they disseminate these videos to farmers using projectors to improve the adoption of agricultural practices. Digital Green’s web-based data tracker (CoCo) stores the attendance and adoption logs of millions of farmers, the videos screened to them and their demographic information. In our work, we leverage this data for a period of ten years between 2010-2020 across five states in India where Digital Green is most active and use it to conduct a holistic evaluation of the ICT. First, we find disparities in the adoption rates of farmers, following which we use statistical tests to identify the different factors that lead to these disparities as well as gender-based inequalities. We find that farmers with higher adoption rates adopt videos of shorter duration and belong to smaller villages. Second, to provide assistance to farmers facing challenges, we model the adoption of practices from a video as a prediction problem and experiment with different model architectures. Our classifier achieves accuracies ranging from 79% to 90% across the five states, demonstrating its potential for assisting future ethnographic investigations. Third, we use SHAP values in conjunction with our model for explaining the impact of various network, content and demographic features on adoption. Our research finds that farmers greatly benefit from past adopters of a video from their group and village. We also discover that videos with a low content-specificity benefit some farmers more than others. Next, we highlight the implications of our findings by translating them into recommendations for providing focused assistance, community building, video screening, revisiting participatory approach and mitigating inequalities. Lastly, we conclude with a discussion on how our work can assist future investigations into the lived experiences of farmers.

Stack Overflow is a Community Question Answering service that attracts millions of users to seek answers to their questions. Maintaining high-quality content is necessary for relevant question retrieval, question recommendation, and enhancing the user experience. Manually removing low-quality content from the platform is time-consuming and challenging for site moderators. Thus, it is imperative to assess the content quality by automatically detecting and ‘closing’ the low-quality questions. Previous works have explored lexical, community-based, vote-based, and style-based features to detect low-quality questions. These approaches are limited to writing styles, textual, and handcrafted features. However, these features fall short in understanding semantic features and capturing the implicit relationships between tags and questions. In contrast, we propose LQuaD (Low-Quality Question Detection), a multi-tier hybrid framework that, a) incorporates semantic information of questions associated with each post using transformers, b) includes the question and tag information that enables learning via a graph convolutional network. LQuaD outperforms the state-of-the-art methods by a 21% higher F1-score on the dataset of 2.8 million questions. Furthermore, we apply survival analysis which acts as a proactive intervention to reduce the number of questions closed by informing users to take appropriate action. We find that the timeframe between the stages from the question's creation till it gets ‘closed’ varies significantly for tags and different ‘closing’ reasons for these questions.

Hashtag segmentation is the task of breaking a hashtag into its constituent tokens. Hashtags often encode the essence of user-generated posts, along with information like topic and sentiment, which are useful in downstream tasks. Hashtags prioritize brevity and are written in unique ways - transliterating and mixing languages, spelling variations, creative named entities. Benchmark datasets used for the hashtag segmentation task - STAN, BOUN - are small and extracted from a single set of tweets. However, datasets should reflect the variations in writing styles of hashtags and account for domain and language specificity, failing which the results will misrepresent model performance. We argue that model performance should be assessed on a wider variety of hashtags, and datasets should be carefully curated. To this end, we propose HashSet, a dataset comprising of: a) 1.9k manually annotated dataset; b) 3.3M loosely supervised dataset. HashSet dataset is sampled from a different set of tweets when compared to existing datasets and provides an alternate distribution of hashtags to build and validate hashtag segmentation models. We analyze the performance of SOTA models for Hashtag Segmentation, and show that the proposed dataset provides an alternate set of hashtags to train and assess models.

Abstract Twitter is an extensively used micro-blogging site for publishing user’s views on recent happenings. This wide reachability of messages over large audience poses a threat, as the degree of personally identifiable information disclosed might lead to user regrets. The Tweet-Scan-Post system scans the tweets contextually for sensitive messages. The tweet repository was generated using cyber-keywords for personal, professional and health tweets. The Rules of Sensitivity and Contextuality was defined based on standards established by various national regulatory bodies. The naive sensitivity regression function uses the Bag-of-Words model built from short text messages. The imbalanced classes in dataset result in misclassification with 25% of sensitive and 75% of insensitive tweets. The system opted stacked classification to combat the problem of imbalanced classes. The system initially applied various state-of-art algorithms and predicted 26% of the tweets to be sensitive. The proposed stacked classification approach increased the overall proportion of sensitive tweets to 35%. The system contributes a vocabulary set of 201 Sensitive Privacy Keyword using the boosting approach for three tweet categories. Finally, the system formulates a sensitivity scaling called TSP’s Tweet Sensitivity Scale based on Senti-Cyber features composed of Sensitive Privacy Keywords, Cyber-keywords with Non-Sensitive Privacy Keywords and Non-Cyber-keywords to detect the degree of disclosed sensitive information.

Comparing two bodies of text and detecting words with significant lexical semantic shift between them is an important part of digital humanities. Traditional approaches have relied on aligning the different embeddings using the Orthogonal Procrustes problem in the Euclidean space. This study presents a geometric framework that leverages smooth Riemannian manifolds for corpus-specific orthogonal rotations and a corpus-independent scaling metric to project the different vector spaces into a shared latent space. This enables us to capture any affine relationship between the embedding spaces while utilising the rich geometry of smooth manifolds.

Twitter has emerged as a prominent social media platform for activism and counterpublic narratives. The counterpublics leverage hashtags to build a diverse support network and share content on a global platform that counters the dominant narrative. This paper applies the framework of connective action on the counter-narrative campaign over the cause of death of #SushantSinghRajput. We combine descriptive network, modularity, and hashtag based topical analysis to identify three major mechanisms underlying the campaign: generative role taking, hashtag-based narratives and formation of alignment network towards a common cause. Using the case study of #SushantSinghRajput, we highlight how connective action framework can be used to identify different strategies adopted by counterpublics for the emergence of connective action.

Social media has grown exponentially in a short period, coming to the forefront of communications and online interactions. Despite their rapid growth, social media platforms have been unable to scale to different languages globally and remain inaccessible to many. In this paper, we characterize Koo, a multilingual micro-blogging site that rose in popularity in 2021, as an Indian alternative to Twitter. We collected a dataset of 4.07 million users, 163.12 million follower-following relationships, and their content and activity across 12 languages. We study the user demographic along the lines of language, location, gender, and profession. The prominent presence of Indian languages in the discourse on Koo indicates the platform's success in promoting regional languages. We observe Koo's follower-following network to be much denser than Twitter's, comprising of closely-knit linguistic communities. An N-gram analysis of posts on Koo shows a #KooVsTwitter rhetoric, revealing the debate comparing the two platforms. Our characterization highlights the dynamics of the multilingual social network and its diverse Indian user base.

The paradigm shift in the consumption of news via online platforms has cultivated the growth of digital journalism. Contrary to traditional media, lowering entry barriers and enabling everyone to be part of content creation have disabled the concept of centralized gatekeeping in digital journalism. This in turn has triggered the production of fake news. Current studies have made a significant effort towards multimodal fake news detection with less emphasis on exploring the discordance between the different multimedia present in a news article. We hypothesize that fabrication of either modality will lead to dissonance between the modalities, and resulting in misrepresented, misinterpreted and misleading news. In this paper, we inspect the authenticity of news coming from online media outlets by exploiting relationship (discordance) between the textual and multiple visual cues. We develop an inter-modality discordance based fake news detection framework to achieve the goal. The modal-specific discriminative features are learned, employing the cross-entropy loss and a modified version of contrastive loss that explores the inter-modality discordance. To the best of our knowledge, this is the first work that leverages information from different components of the news article (i.e., headline, body, and multiple images) for multimodal fake news detection. We conduct extensive experiments on the real-world datasets to show that our approach outperforms the state-of-the-art by an average F1-score of 6.3%.

In recent years, there has been a substantial rise in the consumption of news via online platforms. The ease of publication and lack of editorial rigour in some of these platforms have further led to the proliferation of fake news. In this paper, we study the problem of detecting fake news on the FakeNewsNet repository, a collection of full length articles along with associated images. We present SpotFake+, a multimodal approach that leverages transfer learning to capture semantic and contextual information from the news articles and its associated images and achieves the better accuracy for fake news detection. To the best of our knowledge, this is the first work that performs a multimodal approach for fake news detection on a dataset that consists of full length articles. It outperforms the performance shown by both single modality and multiple-modality models. We also release the pretrained model for the benefit of the community.

While extensive popularity of online social media platforms has made information dissemination faster, it has also resulted in widespread online abuse of different types like hate speech, offensive language, sexist and racist opinions, etc.Detection and curtailment of such abusive content is critical for avoiding its psychological impact on victim communities, and thereby preventing hate crimes.Previous works have focused on classifying user posts into various forms of abusive behavior.But there has hardly been any focus on estimating the severity of abuse and the target.In this paper, we present a first of the kind dataset with 7,601 posts from Gab 1 which looks at online abuse from the perspective of presence of abuse, severity and target of abusive behavior.We also propose a system to address these tasks, obtaining an accuracy of ∼80% for abuse presence, ∼82% for abuse target prediction, and ∼65% for abuse severity prediction.

In multilingual societies like the Indian subcontinent, use of code-switched languages is much popular and convenient for the users. In this paper, we study offense and abuse detection in the code-switched pair of Hindi and English (i.e, Hinglish), the pair that is the most spoken. The task is made difficult due to non-fixed grammar, vocabulary, semantics and spellings of Hinglish language. We apply transfer learning and make a LSTM based model for hate speech classification. This model surpasses the performance shown by the current best models to establish itself as the state-of-the-art in the unexplored domain of Hinglish offensive text classification. We also release our model and the embeddings trained for research purposes.

In the era of MOOCs, online exams are taken by millions of candidates, where scoring short answers is an integral part. It becomes intractable to evaluate them by human graders. Thus, a generic automated system capable of grading these responses should be designed and deployed. In this paper, we present a fast, scalable, and accurate approach towards automated Short Answer Scoring (SAS). We propose and explain the design and development of a system for SAS, namely AutoSAS. Given a question along with its graded samples, AutoSAS can learn to grade that prompt successfully. This paper further lays down the features such as lexical diversity, Word2Vec, prompt, and content overlap that plays a pivotal role in building our proposed model. We also present a methodology for indicating the factors responsible for scoring an answer. The trained model is evaluated on an extensively used public dataset, namely Automated Student Assessment Prize Short Answer Scoring (ASAP-SAS). AutoSAS shows state-of-the-art performance and achieves better results by over 8% in some of the question prompts as measured by Quadratic Weighted Kappa (QWK), showing performance comparable to humans.

Massively Multiplayer Online Role-Playing Games (MMORPGs) are persistent virtual environments where millions of players interact in an online manner. We study the problem of player churn and social contagion using MMORPG game logs by analyzing the impact of a node's churn behavior on its immediate neighborhood or group. The two key research questions in this paper are - When an active node, ego, becomes dormant, what is the impact on the activity behavior of ego's immediate neighbor, alter, 1) based on ego's characteristics and ego's relationship with alter and 2) based on the activity behavior of alter's remaining neighbors. We use a supervised learning framework to study the impact of player churn and social contagion. Experimental results show that the classification models perform substantially better than random for both the research problems. Finally, we use a data-driven approach to propose a player typology based on degree of socialization and analyze churn behavior among these player types. Experimental results show that the loner player type is much more likely to churn than the socializer player types and as the degree of socialization decreases among socializers, the propensity to churn increases.

Social dynamics are based on human needs for trust, support, resource sharing, irrespective of whether they operate in real life or in a virtual setting. Massively multiplayer online role-playing games (MMORPGS) serve as enablers of leisurely social activity and are important tools for social interactions. Past research has shown that socially dense gaming environments like MMORPGs can be used to study important social phenomena, which may operate in real life, too. We describe the process of social exploration to entail the following components 1) finding the balance between personal and social time 2) making choice between a large number of weak ties or few strong social ties. 3) finding a social group. In general, these are the major determinants of an individual's social life. This paper looks into the phenomenon of social exploration in an activity based online social environment. We study this process through the lens of the following research questions, 1) What are the different social behavior types? 2) Is there a change in a player's social behavior over time? 3) Are certain social behaviors more stable than the others? 4) Can longitudinal research of player behavior help shed light on the social dynamics and processes in the network? We use an unsupervised machine learning approach to come up with 4 different social behavior types - Lone Wolf, Pack Wolf of Small Pack, Pack Wolf of a Large Pack and Social Butterfly. The types represent the degree of socialization of players in the game. Our research reveals that social behaviors change with time. While lone wolf and pack wolf of small pack are more stable social behaviors, pack wolf of large pack and social butterflies are more transient. We also observe that players progressively move from large groups with weak social ties to settle in small groups with stronger ties.

Livestreaming platforms enable content producers, or streamers, to broadcast creative content to a potentially large viewer base. Chatrooms form an integral part of such platforms, enabling viewers to interact both with the streamer, and amongst themselves. Streams with high engagement (many viewers and active chatters) are typically considered engaging, and often promoted to end users by means of recommendation algorithms, and exposed to better monetization opportunities via revenue share from platform advertising, viewer donations, and third-party sponsorships. Given such incentives, some streamers make use of fraudulent means to increase perceived engagement by simulating chatter via fake "chatbots" which can be purchased from shady online marketplaces. This inauthentic engagement can negatively influence recommendation, hurt streamer and viewer trust in the platform, and harm monetization for honest streamers. In this paper, we tackle the novel problem of automating detection of chatbots on livestreaming platforms. To this end, we first formalize the livestreaming chatbot detection problem and characterize differences between botted and genuine chatter behavior observed from a real-world livestreaming chatter dataset collected from Twitch.tv. We then propose SHERLOCK, which posits a two-stage approach of detecting chatbotted streams, and subsequently detecting the constituent chatbots. Finally, we demonstrate effectiveness on both real and synthetic data: to this end, we propose a novel strategy for collecting labeled, synthetic chatter dataset (typically unavailable) from such platforms, enabling evaluation of proposed detection approaches against chatbot behaviors with varying signatures. Our approach achieves .97 precision/recall on the real-world dataset, and .80+ F1 scores across most simulated attack settings.

Road safety mapping using satellite images is a cost-effective but a challenging problem for smart city planning. The scarcity of labeled data, misalignment and ambiguity makes it hard to learn efficient embeddings in order to classify between safe and dangerous road segments. In this paper, we address the challenges using a region guided attention network. In our model, we extract global features from a base network and augment it with local features obtained using the region guided attention network. In addition, we perform domain adaptation for unlabeled target data. In order to bridge the gap between safe samples and dangerous samples from source and target respectively, we propose a loss function based on within and between class covariance matrices. We conduct experiments on a public dataset of London to show that the algorithm achieves significant results with the classification accuracy of 86.21%. We obtain an increase of 4% accuracy for NYC using domain adaptation network.

Deep neural networks (DNNs) are vulnerable to malicious inputs crafted by an adversary to produce erroneous outputs. Works on securing neural networks against adversarial examples achieve high empirical robustness on simple datasets such as MNIST. However, these techniques are inadequate when empirically tested on complex data sets such as CIFAR-10 and SVHN. Further, existing techniques are designed to target specific attacks and fail to generalize across attacks. We propose Adversarial Model Cascades (AMC) as a way to tackle the above inadequacies. Our approach trains a cascade of models sequentially where each model is optimized to be robust towards a mixture of multiple attacks. Ultimately, it yields a single model which is secure against a wide range of attacks; namely FGSM, Elastic, Virtual Adversarial Perturbations and Madry. On an average, AMC increases the model's empirical robustness against various attacks simultaneously, by a significant margin (of 6.225% for MNIST, 5.075% for SVHN and 2.65% for CIFAR-10 ). At the same time, the model's performance on non-adversarial inputs is comparable to the state-of-the-art models.

In recent years, there has been an explosion in the number of memes being created and circulated in online social networks. Despite their rapidly increasing impact on how we communicate online, meme images are virtually inaccessible to the visually impaired users. Existing automated assistive systems that were primarily devised for natural photos in social media, overlook the specific fine-grained visual details in meme images. In this paper, we concentrate on describing one such prominent visual detail: the meme face emotion. We propose a novel automated method that enables visually impaired social media users to understand and appreciate meme face emotions with the help of rich textual captions. We first collect a challenging dataset of meme face emotion captions to support future research in face emotion understanding. We design a two-stage approach that significantly outperforms baseline approaches across all the standard captioning metrics and also generates richer discriminative captions. By validating our solution with the help of visually impaired social media users, we show that our emotion captions enable them to understand and appreciate one of the most popular classes of meme images encountered on the Internet for the first time. Code, data, and models are publicly available.

Stack Overflow, a Q&A site on programming, awards reputation points and badges (game elements) to users on performing various actions. Situating our work in Digital Signaling Theory, we investigate the role of these game elements in characterizing social qualities (specifically, popularity and impact) of its users. We operationalize these attributes using common metrics and apply statistical modeling to empirically quantify and validate the strength of these signals. Our results are based on a rich dataset of 3,831,147 users and their activities spanning nearly a decade since the site's inception in 2008. We present evidence that certain non-trivial badges, reputation scores and age of the user on the site positively correlate with popularity and impact. Further, we find that the presence of costly to earn and hard to observe signals qualitatively differentiates highly impactful users from highly popular users.

User timelines in Online Social Media (OSM) remains filled with a significant amount of information received from followees. Given that content posted by followee is not under user's control, this information may not always be relevant. If there is large presence of not so relevant content, then a user may end up overlooking relevant content, which is undesirable. To address this issue, in the first part of our work, we propose suitable metrics to characterize the user-followee relationship. We find that most of the users choose their followees primarily due to the content that they post (content-conscious behavior, measured by content similarity scores). For a small number of followees, a high degree of social engagement (likes and shares) irrespective of the content posted by them is observed (user-conscious behavior, measured by user affinity scores). We evaluate our proposed approach on 26,516 followees across 100 random users on Twitter who have cumulatively posted 234,403 tweets. We find that on average for 60 % of their followees, users exhibit very low degree of content similarity and social engagement. These findings motivate the second part of our work, where we develop a Followee Management Nudge (FMN) through a browser extension (plugin) that helps users remain more informed about their relationship with each of their followees. In particular, the FMN nudges a user with a list of followees with whom they have least (or never) engaged in the past and also exhibit very low similarity in terms of content, thereby helping a user to make an informed decision (say by unfollowing some of these followees). Results from a preliminary controlled lab study show that 62.5 % of participants find the nudge to be quite useful.

In this work, we take a step towards understanding and defending against spying browser extensions. These are extensions repurposed to capture online activities of a user and communicate the collected sensitive information to a third-party domain. We conduct an empirical study of such extensions on the Chrome Web Store. First, we present an in-depth analysis of the spying behavior of these extensions. We observe that these extensions steal a variety of sensitive user information, such as the complete browsing history (e.g., the sequence of web traversals), online social network (OSN) access tokens, IP address, and geolocation. Second, we investigate the potential for automatically detecting spying extensions by applying machine learning schemes. We show that using a Recurrent Neural Network (RNN), the sequence of browser API calls made by an extension can be a robust feature, outperforming hand-crafted features (used in prior work on malicious extensions) to detect spying extensions. Our RNN based detection scheme achieves a high precision (90.02%) and recall (93.31%) in detecting spying extensions.

Ransomware is a malware which is spreading largely around the world and imposing serious threats to information assets and victimizes internet users by hijacking and encrypting their files, and then demanding payment to get back access to the files. Seeking system vulnerabilities, ransomware tries to seize control over the victim's files and system, until the victim agrees to the attacker's demands. In this paper, we have extended the work by focusing on Indicator of Compromises (IOCs) for ransomware using Cuckoo Sandbox. These IOCs are used to set the base for the analysis and classification of new ransomware into their respective class. Further, working for the implementation of the proposed idea, we have automated the system using machine learning algorithms for the classification of ransomware variants in the real time environment. To the best of our knowledge, this is the first real time, comprehensive and usable methodology to classify the ransomware in their respective classes.

Internet has brought about a tremendous increase in content of all forms and, in that, video content constitutes the major backbone of the total content being published as well as watched. Thus it becomes imperative for video recommendation engines to look for novel and innovative ways to recommend the newly added videos to their users. However, the problem with new videos is that they lack any sort of metadata and user interaction so as to be able to rate the videos for the consumers. To this effect, this paper introduces the several techniques we develop for the Content Based Video Relevance Prediction (CBVRP). We employ different architectures on the CBVRP dataset to make use of the provided frame and video level features and generate predictions of videos that are similar to the other videos. We also implement several ensemble strategies to explore complementarity between both the types of provided features. The obtained results are encouraging and will impel the boundaries of research for multimedia based video recommendation systems.

Cybercriminals have leveraged the popularity of a large user base available on Online Social Networks~(OSNs) to spread spam campaigns by propagating phishing URLs, attaching malicious contents, etc. However, another kind of spam attacks using phone numbers has recently become prevalent on OSNs, where spammers advertise phone numbers to attract users» attention and convince them to make a call to these phone numbers. The dynamics of phone number based spam is different from URL-based spam due to an inherent trust associated with a phone number. While previous work has proposed strategies to mitigate URL-based spam attacks, phone number based spam attacks have received less attention.

Extensive research has been conducted to identify, analyze and measure popular topics and public sentiment on Online Social Networks (OSNs) through text, especially during crisis events. However, little work has been done to understand such events through pictures posted on these networks. Given the potential of visual content for influencing users' thoughts and emotions, we perform a large-scale analysis to study and compare popular themes and sentiment across images and textual content posted on Facebook during the terror attacks that took place in Paris in 2015. We propose a generalizable and highly automated 3-tier pipeline which utilizes state-of-the-art computer vision techniques to extract high-level human understandable image descriptors. We used these descriptors to associate themes and sentiment with images, and analyzed over 57,000 images related to the Paris Attacks. We discovered multiple visual themes which were popular in images, but were not identifiable through text. We also uncovered instances of misinformation and false flag (conspiracy) theories among popular image themes, which were not prominent in user-generated textual content. Further, our analysis revealed that while textual content posted after the attacks reflected negative sentiment, images inspired positive sentiment. These findings suggest that large-scale mining of images posted on OSNs during crisis, and other news-making events can significantly augment textual content to understand such events.

Identifying medical persona from a social media post is of paramount importance for drug marketing and pharmacovigilance. In this work, we propose multiple approaches to infer the medical persona associated with a social media post. We pose this as a supervised multi-label text classification problem. The main challenge is to identify the hidden cues in a post that are indicative of a particular persona. We first propose a large set of manually engineered features for this task. Further, we propose multiple neural network based architectures to extract useful features from these posts using pre-trained word embeddings. Our experiments on thousands of blogs and tweets show that the proposed approach results in 7% and 5% gain in F-measure over manual feature engineering based approach for blogs and tweets respectively.

The Islamic State of Iraq and Syria (ISIS) is a Salafi jihadist militant group that has made extensive use of online social media platforms to promulgate its ideologies and evoke many individuals to support the organization. The psycho-sociological background of an individual plays a crucial role in determining his/her vulnerability of being lured into joining the organisation and indulge in terrorist activities, since his/her behavior largely depends on the society s/he was brought up in. Here, we analyse five sociological aspects -- personality, values & ethics, optimism/pessimism, age and gender to understand the psycho-sociological vulnerability of individuals over Twitter. Experimental results suggest that psycho-sociological aspects indeed act as foundation to discover and differentiate between prominent and unobtrusive users in Twitter.

Psychological distress in the form of depression, anxiety and other mental health challenges among college students is a growing health concern. Dearth of accurate, continuous, and multi-campus data on mental well-being presents significant challenges to intervention and mitigation efforts in college campuses. We examine the potential of social media as a new "barometer" for quantifying the mental well-being of college populations. Utilizing student-contributed data in Reddit communities of over 100 universities, we first build and evaluate a transfer learning based classification approach that can detect mental health expressions with 97% accuracy. Thereafter, we propose a robust campus-specific Mental Well-being Index: MWI. We find that MWI is able to reveal meaningful temporal patterns of mental well-being in campuses, and to assess how their expressions relate to university attributes like size, academic prestige, and student demographics. We discuss the implications of our work for improving counselor efforts, and in the design of tools that can enable better assessment of the mental health climate of college campuses.

Social media platforms have obtained substantial interest of police to connect with residents. This has encouraged residents to report day-to-day law and order concerns such as traffic congestion, missing people, and harassment by cops on these platforms. In this paper, we study day-to-day concerns shared by residents on social media and police response to such concerns. Based on the input of police experts, we define concerns that require police response and attention, as a serviceable request. We provide insights on six textual attributes that can identify serviceable posts. We find such posts are marked by high negative emotions, more factual, and objective content such as location and time of incidences. We show that police response time varies depending upon the kind of serviceable requests. Our work explores a series of statistical models to predict serviceable posts and its different types. We conclude the paper, discussing the implication of our findings on police practices and design needs for possible technological interventions. These technological interventions will help increase the interactions between police and residents and thereby increasing the well-being and safety of society.

In this paper we analyze Facebook's Free Basics program, which provides free Internet access to a restricted set of web services. As the program grows to 60+ developing countries, an independent and data-driven audit of its scope and outreach is highly relevant to the ICTD community.

The data generated on social media sites continues to grow at an increasing rate with more than 36% of tweets containing images making the dominance of multimedia content evidently visible. This massive user generated content has become a reflection of world events. In order to enhance the ability and effectiveness to consume this plethora of data, summarization of these events is needed. However, very few studies have exploited the images attached with social media events to summarize them using "mid-level visual elements". These are the entities which are both representative and discriminative to the target dataset besides being human-readable and hence more informative.

Existing CAPTCHA solutions are a major source of user frustration on the Internet today, frequently forcing companies to lose customers and business. Game CAPTCHAs are a promising approach which may make CAPTCHA solving a fun activity for the user. One category of such CAPTCHAs – called Dynamic Cognitive Game (DCG) CAPTCHA – challenges the user to perform a game-like cognitive (or recognition) task interacting with a series of dynamic images. Specifically, it takes the form of many objects floating around within the images, and the user’s task is to match the objects corresponding to specific target(s), and drag/drop them to the target region(s). In this paper, we pursue a comprehensive analysis of DCG CAPTCHAs. We design and implement such CAPTCHAs, and dissect them across four broad but overlapping dimensions: (1) usability, (2) fully automated attacks, (3) human-solving relay attacks, and (4) hybrid attacks that combine the strengths of automated and relay attacks. Our study shows that DCG CAPTCHAs are highly usable, even on mobile devices and offer some resilience to relay attacks, but they are vulnerable to our proposed automated and hybrid attacks.

Facebook is the world's largest Online Social Network, having more than 1 billion users. Like most other social networks, Facebook is home to various categories of hostile entities who abuse the platform by posting malicious content. In this paper, we identify and characterize Facebook pages that engage in spreading URLs pointing to malicious domains. We revisit the scope and definition of what is deemed as “malicious” in the modern day Internet, and identify 627 pages publishing untrustworthy information, misleading content, adult and child unsafe content, scams, etc. Our findings revealed that at least 8% of all malicious pages were dedicated to promote a single malicious domain. Studying the temporal posting activity of pages revealed that malicious pages were 1.4 times more active daily than benign pages. We further identified collusive behavior within a set of malicious pages spreading adult and pornographic content. Finally, we attempted to automate the process of detecting malicious Facebook pages by training multiple supervised learning algorithms on our dataset. Artificial neural networks trained on a fixed sized bag-of-words performed the best and achieved an accuracy of 84.13%.

Increasing cyber-security presents an ongoing challenge to security professionals. Research continuously suggests that online users are a weak link in information security. This research explores the relationship between cyber-security and cultural, personality and demographic variables.

YouTube draws large number of users who contribute actively by uploading videos or commenting on existing videos. However, being a crowd sourced and large content pushed onto it, there is limited control over the content. This makes malicious users push content (videos and comments) which is inappropriate (unsafe), particularly when such content is placed around cartoon videos which are typically watched by kids. In this paper, we focus on presence of unsafe content for children and users who promote it. For detection of child unsafe content and its promoters, we perform two approaches, one based on supervised classification which uses an extensive set of video-level, user-level and comment-level features and another based Convolutional Neural Network using video frames. Detection accuracy of 85.7% is achieved which can be leveraged to build a system to provide a safe YouTube experience for kids. Through detailed characterization studies, we are able to successfully conclude that unsafe content promoters are less popular and engage less as compared with other users. Finally, using a network of unsafe content promoters and other users based on their engagements (likes, subscription and playlist addition) and other factors, we find that unsafe content is present very close to safe content and unsafe content promoters form very close knit communities with other users, thereby further increasing the likelihood of a child getting getting exposed to unsafe content.

Organizations measure their social audience based on the number of users, fans, and followers on social media. Every social media platform has its user identity and a single user is present across varied platforms. Due to the disconnected user profiles, identifying duplicate users across media is non-trivial. There is a need to create a complete view of a user for various applications such as targeting and user profile construction. This view is not easily available due to the individual identities. In this work, we explore the feature space across social media that can be leveraged for intelligent user identity aggregation. Further, we present a two-phased unified identity creation process using our feature analysis, unsupervised candidate selection, and supervised user matching algorithms on four different social networks.

Internet users and businesses are increasingly using online social networks (OSN) to drive audience traffic and increase their popularity. In order to boost social presence, OSN users need to increase the visibility and reach of their online profile, like - Facebook likes, Twitter followers, Instagram comments and Yelp reviews. For example, an increase in Twitter followers not only improves the audience reach of the user but also boosts the perceived social reputation and popularity. This has led to a scope for an underground market that provides followers, likes, comments, etc. via a network of fraudulent and compromised accounts and various collusion techniques. In this paper, we landscape the underground markets that provide Twitter followers by studying their basic building blocks - merchants, customers and phony followers. We charecterize the services provided by merchants to understand their operational structure and market hierarchy. Twitter underground markets can operationalize using a premium monetary scheme or other incentivized freemium schemes. We find out that freemium market has an oligopoly structure with few merchants being the market leaders. We also show that merchant popularity does not have any correlation with the quality of service provided by the merchant to its customers. Our findings also shed light on the characteristics and quality of market customers and the phony followers provided by underground market. We draw comparison between legitimate users and phony followers, and find out key identifiers to separate such users. With the help of these differentiating features, we build a supervised learning model to predict suspicious following behaviour with an accuracy of 89.2%.

Online Social Networks (OSNs) witness a rise in user activity whenever a news-making event takes place. Cyber criminals exploit this spur in user-engagement levels to spread malicious content that compromises system reputation, causes financial losses and degrades user experience. In this paper, we characterized a dataset of 4.4 million public posts generated on Facebook during 17 news-making events (natural calamities, terror attacks, etc.) and identified 11,217 malicious posts containing URLs. We found that most of the malicious content which is currently evading Facebook's detection techniques originated from third party and web applications, while more than half of all legitimate content originated from mobile applications. We also observed greater participation of Facebook pages in generating malicious content as compared to legitimate content. We proposed an extensive feature set based on entity profile, textual content, metadata, and URL features to automatically identify malicious content on Facebook in real time. This feature set was used to train multiple machine learning models and achieved an accuracy of 86.9%. We performed experiments to show that past techniques for spam campaign detection identified less than half the number of malicious posts as compared to our model. This model was used to create a REST API and a browser plug-in to identify malicious Facebook posts in real time.

Existing captcha solutions on the Internet are a major source of user frustration. Game captchas are an interesting and, to date, little-studied approach claiming to make captcha solving a fun activity for the users. One broad form of such captchas -- called Dynamic Cognitive Game (DCG) captchas -- challenge the user to perform a game-like cognitive task interacting with a series of dynamic images. We pursue a comprehensive analysis of a representative category of DCG captchas. We formalize, design and implement such captchas, and dissect them across: (1) fully automated attacks, (2) human-solver relay attacks, and (3) usability. Our results suggest that the studied DCG captchas exhibit high usability and, unlike other known captchas, offer some resistance to relay attacks, but they are also vulnerable to our novel dictionary-based automated attack.

Mobile-based branchless banking has become a key mechanism for enabling financial inclusion in the developing world. A key component of all branchless banking systems is a mechanism to provide receipts to users after each transaction as evidence for successful transaction completion. In this paper, we present results from a field study that explores user perceptions of different receipt delivery mechanisms in the context of a branchless banking system in India. Our study shows that users have an affinity for paper receipts: despite the provision of an SMS receipt functionality by the system developers and their discouragement of the use of paper, users have pro-actively initiated a practice of issuing and accepting paper receipts. Several users are aware of the security limitations of paper receipts but continue to use them because of their usability benefits. We conclude with design recommendations for receipt delivery systems in branchless banking.

With the growing popularity and usage of online social media services, people now have accounts (some times several) on multiple and diverse services like Facebook, Linked In, Twitter and You Tube. Publicly available information can be used to create a digital footprint of any user using these social media services. Generating such digital footprints can be very useful for personalization, profile management, detecting malicious behavior of users. A very important application of analyzing users' online digital footprints is to protect users from potential privacy and security risks arising from the huge publicly available user information. We extracted information about user identities on different social networks through Social Graph API, Friend Feed, and Profilactic, we collated our own dataset to create the digital footprints of the users. We used username, display name, description, location, profile image, and number of connections to generate the digital footprints of the user. We applied context specific techniques (e.g. Jaro Winkler similarity, Word net based ontologies) to measure the similarity of the user profiles on different social networks. We specifically focused on Twitter and Linked In. In this paper, we present the analysis and results from applying automated classifiers for disambiguating profiles belonging to the same user from different social networks. User ID and Name were found to be the most discriminative features for disambiguating user profiles. Using the most promising set of features and similarity metrics, we achieved accuracy, precision and recall of 98%, 99%, and 96%, respectively.

In the last few years, the increasing interest in location-based services (LBS) has favored the introduction of geo-referenced information in various Web 2.0 applications, as well as the rise of location-based social networks (LBSN). Foursquare, one of the most popular LBSNs, gives incentives to users who visit (check in) specific places (venues) by means of, for instance, mayorships to frequent visitors. Moreover, users may leave tips at specific venues as well as mark previous tips as done in sign of agreement. Unlike check ins, which are shared only with friends, the lists of mayorships, tips and dones of a user are publicly available to everyone, thus raising concerns about disclosure of the user's movement patterns and interests. We analyze how users explore these publicly available features, and their potential as sources of information leakage. Specifically, we characterize the use of mayorships, tips and dones in Foursquare based on a dataset with around 13 million users. We also analyze whether it is possible to easily infer the home city (state and country) of a user from these publicly available information. Our results indicate that one can easily infer the home city of around 78% of the analyzed users within 50 kilometers.

Short Message Service (SMS) is one of the most frequently used services in the mobile phones, next to calls. In developing countries like India, SMS is the cheapest mode of communication. The advantage of this fact is exploited by the advertising companies to reach masses. The unsolicited SMS messages (a.k.a. spam SMS) generates notifications, thus consuming precious user attention. To formulate spam SMS problem and understand user's needs and preceptions, we conducted an online survey with 458 participants in different cities of India. Most of the survey participants admitted that they are quite annoyed with burst of SMS spams and in-effectiveness of regulatory solutions. However, some participants reported that, they do get useful information from spam SMSes sometime(e.g. discounts at a popular food joint). In this paper, we present design and implementation of a user-centric spam SMS filtering application i.e. SMSAssassin that uses content based machine learning techniques with user generated features to filter unwanted SMSes and reduces the burden of notifications for a mobile user.

In this paper we propose and evaluate Marasim, a novel Jigsaw based graphical authentication mechanism using tagging. Marasim is aimed at achieving the security of random images with the memorability of personal images. Our scheme relies on the human ability to remember a personal image and later recognize the alternate visual representations (images) of the concepts occurred in the image. These concepts are retrieved from the tags assigned to the image. We illustrate how a Jigsaw based approach helps to create a portfolio of system-chosen random images to be used for authentication. The paper describes the complete design of Marasim along with the empirical studies of Marasim that provide evidences of increased memorability. Results show that 93% of all participants succeeded in the authentication tests using Marasim after three months while 71% succeeded in authentication tests using Marasim after nine months. Our findings indicate that Marasim has potential applications, especially where text input is hard (e.g., PDAs or ATMs), or in situations where passwords are infrequently used (e.g., web site passwords).

Several studies ranked security and privacy to be major areas of concern and impediments of cloud adoption for companies, but none have looked into end-users' attitudes and practices. Not much is known about consumers' privacy beliefs and expectations for cloud storage, such as web-mail, document and photo sharing platforms, or about users' awareness of contractual terms and conditions. We conducted 36 in-depth interviews in Switzerland and India (two countries with different privacy perceptions and expectations); and followed up with an online survey with 402 participants in both countries. We study users' privacy attitudes and beliefs regarding their use of cloud storage systems. Our results show that privacy requirements for consumer cloud storage differ from those of companies. Users are less concerned about some issues, such as guaranteed deletion of data, country of storage and storage outsourcing, but are uncertain about using cloud storage. Our results further show that end-users consider the Internet intrinsically insecure and prefer local storage for sensitive data over cloud storage. However, users desire better security and are ready to pay for services that provide strong privacy guarantees. Participants had misconceptions about the rights and guarantees their cloud storage providers offers. For example, users believed that their provider is liable in case of data loss, does not have the right to view and modify user data, and cannot disable user accounts. Finally, our results show that cultural differences greatly influence user attitudes and beliefs, such as their willingness to store sensitive data in the cloud and their acceptance that law enforcement agencies monitor user accounts. We believe that these observations can help in improving users privacy in cloud storage systems.

In this paper we present the results of a roleplay survey instrument administered to 1001 online survey respondents to study both the relationship between demographics and phishing susceptibility and the effectiveness of several anti-phishing educational materials. Our results suggest that women are more susceptible than men to phishing and participants between the ages of 18 and 25 are more susceptible to phishing than other age groups. We explain these demographic factors through a mediation analysis. Educational materials reduced users' tendency to enter information into phishing webpages by 40% percent; however, some of the educational materials we tested also slightly decreased participants' tendency to click on legitimate links.

Recent years have seen a proliferation of secure device pairing methods that try to improve both the usability and security of today's de-facto standard -- PIN-based authentication. Evaluating such improvements is difficult. Most comparative laboratory studies have so far mainly focused on completeness, trying to find the single best method among the dozens of proposed approaches -- one that is both rated the most usable by test subjects, and which provides the most robust security guarantees. This search for the "best" pairing method, however, fails to take into account the variety of situations in which such pairing protocols may be used in real life. The comparative study reported here, therefore, explicitly situates pairing tasks in a number of more realistic situations. Our results indicate that people do not always use the easiest or most popular method -- they instead prefer different methods in different situations, based on the sensitivity of data involved, their time constraints, and the social conventions appropriate for a particular place and setting. Our study also provides qualitative data on factors influencing the perceived security of a particular method, the users' mental models surrounding security of a method, and their security needs.

Phishing attacks, in which criminals lure Internet users to Web sites that spoof legitimate Web sites, are occurring with increasing frequency and are causing considerable harm to victims. While a great deal of effort has been devoted to solving the phishing problem by prevention and detection of phishing emails and phishing Web sites, little research has been done in the area of training users to recognize those attacks. Our research focuses on educating users about phishing and helping them make better trust decisions. We identified a number of challenges for end-user security education in general and anti-phishing education in particular: users are not motivated to learn about security; for most users, security is a secondary task; it is difficult to teach people to identify security threats without also increasing their tendency to misjudge nonthreats as threats. Keeping these challenges in mind, we developed an email-based anti-phishing education system called “PhishGuru” and an online game called “Anti-Phishing Phil” that teaches users how to use cues in URLs to avoid falling for phishing attacks. We applied learning science instructional principles in the design of PhishGuru and Anti-Phishing Phil. In this article we present the results of PhishGuru and Anti-Phishing Phil user studies that demonstrate the effectiveness of these tools. Our results suggest that, while automated detection systems should be used as the first line of defense against phishing attacks, user education offers a complementary approach to help people better recognize fraudulent emails and websites.

PhishGuru is an embedded training system that teaches users to avoid falling for phishing attacks by delivering a training message when the user clicks on the URL in a simulated phishing email. In previous lab and real-world experiments, we validated the effectiveness of this approach. Here, we extend our previous work with a 515-participant, real-world study in which we focus on long-term retention and the effect of two training messages. We also investigate demographic factors that influence training and general phishing susceptibility. Results of this study show that (1) users trained with PhishGuru retain knowledge even after 28 days; (2) adding a second training message to reinforce the original training decreases the likelihood of people giving information to phishing websites; and (3) training does not decrease users' willingness to click on links in legitimate messages. We found no significant difference between males and females in the tendency to fall for phishing emails both before and after the training. We found that participants in the 18--25 age group were consistently more vulnerable to phishing attacks on all days of the study than older participants. Finally, our exit survey results indicate that most participants enjoyed receiving training during their normal use of email.

Phishing attacks, in which criminals lure Internet users to websites that impersonate legitimate sites, are occurring with increasing frequency and are causing considerable harm to victims. In this paper we describe the design and evaluation of an embedded training email system that teaches people about phishing during their normal use of email. We conducted lab experiments contrasting the effectiveness of standard security notices about phishing with two embedded training designs we developed. We found that embedded training works better than the current practice of sending security notices. We also derived sound design principles for embedded training systems.

In this paper we describe the design and evaluation of Anti-Phishing Phil, an online game that teaches users good habits to help them avoid phishing attacks. We used learning science principles to design and iteratively refine the game. We evaluated the game through a user study: participants were tested on their ability to identify fraudulent web sites before and after spending 15 minutes engaged in one of three anti-phishing training activities (playing the game, reading an anti-phishing tutorial we created based on the game, or reading existing online training materials). We found that the participants who played the game were better able to identify fraudulent web sites compared to the participants in other conditions. We attribute these effects to both the content of the training messages presented in the game as well as the presentation of these materials in an interactive game format. Our results confirm that games can be an effective way of educating people about phishing and other security attacks.

Trust is an important component of online transactions. The increasing amount and sophistication of spam, phishing, and other semantic attacks increase users' uncertainty about the consequences of their actions and their distrust towards other online parties. In this paper, we highlight some key characteristics of a model that we are developing to represent and compare the online trust decision processes of "expert" and "non-expert" computer users. We also report on preliminary data we are gathering to validate, refine, and apply our model. This research is part of a broader project that aims at developing tools and training modules to help online users make good trust decisions.