WWW 2025 Papers

Recently, the jailbreak attack, which generates adversarial prompts to bypass safety measures and mislead large language models (LLMs) to output harmful answers, has attracted extensive interest due to its potential to reveal the vulnerabilities of LLMs. However, ignoring the exploitation of the characteristics in intention understanding, existing studies could only generate prompts with weak attacking ability, failing to evade defenses (e.g., sensitive word detect) and causing malice(e.g., harmful outputs). Motivated by the mechanism in the psychology of human misjudgment, we propose a dual intention escape (DIE) jailbreak attack framework to generate more stealthy and toxic prompts to deceive LLMs to output harmful content. For stealthiness, inspired by the anchoring effect, we designed the Intention-anchored Malicious Concealment(IMC) module that hides the harmful intention behind a generated anchor intention by the recursive decomposition block and contrary intention nesting block. Since the anchor intention will be received first, the LLMs might pay less attention to the harmful intention and enter response status. For toxicity, we propose the Intention-reinforced Malicious Inducement (IMI) module based on the availability bias mechanism in a progressive malicious prompting approach. Due to the ongoing emergence of statements correlated to harmful intentions, the output content of LLMs will be closer to these more accessible intentions, $\textit{i.e.}$, more toxic. We conducted extensive experiments under black-box settings, supporting that DIE could achieve 100\% ASR-R and 92.9\% ASR-G against GPT3.5-turbo.

With the development of blockchain technology, crypto gambling has gained popularity due to its high level of anonymity. However, similar to traditional casinos, crypto casinos are controlled by a few internal $\textit{Delegatees}$, making it impossible for them to achieve complete transparency and fairness. These delegatees are hidden among $\textit{gamblers}$ and are difficult to identify and distinguish in anonymous and large-scale blockchain transaction networks. This paper proposes an unsupervised dual-stage role identification method to adaptively identify key roles and hidden delegatees in label-sparse crypto casinos. Specifically, inspired by voting-style transaction patterns, we propose a novel voting influence metric for key node identification. This metric is based on one-dimensional structural entropy to capture global dissemination capability. Subsequently, we develop a multi-view graph neural network framework enhanced with two-dimensional global structural entropy minimization and self-supervised contrastive learning to improve the robustness and interpretability of hidden role partitioning. Experiments on real-world cases of the most mainstream blockchains—Ethereum, TRON, and Arbitrum—demonstrate that our proposed method effectively reveals distinct role compositions and collusion patterns, distinguishing between gamblers and delegatees. Our results achieve a higher match with identities confirmed by judicial authorities than existing methods, indicating the effectiveness and generalizability of our approach in enhancing security and regulation oversight.

Recommender systems (RS) are pivotal in managing information overload in modern digital services. A key challenge in RS is efficiently processing vast item pools to deliver highly personalized recommendations under strict latency constraints. Multi-stage cascade ranking addresses this by employing computationally efficient retrieval methods to cover diverse user interests, followed by more precise ranking models to refine the results. In the retrieval stage, multi-channel retrieval is often used to generate distinct item subsets from different candidate generators, leveraging the complementary strengths of these methods to maximize coverage. However, forwarding all retrieved items overwhelms downstream rankers, necessitating truncation. Despite advancements in individual retrieval methods, multi-channel fusion, the process of efficiently merging multi-channel retrieval results, remains underexplored. We are the first to identify and systematically investigate multi-channel fusion in the retrieval stage. Current industry practices often rely on heuristic approaches and manual designs, which often lead to suboptimal performance. Moreover, traditional gradient-based methods like SGD are unsuitable for this task due to the non-differentiable nature of the selection process. In this paper, we explore advanced channel fusion strategies by assigning systematically optimized weights to each channel. We utilize black-box optimization techniques, including the Cross Entropy Method and Bayesian Optimization for global weight optimization, alongside policy gradient-based approaches for personalized merging. Our methods enhance both personalization and flexibility, achieving significant performance improvements across multiple datasets and yielding substantial gains in real-world deployments, offering a scalable solution for optimizing multi-channel fusion in retrieval.

Large language models (LLMs) enhanced with retrieval-augmented generation (RAG) have introduced a new paradigm for web search. However, the limited context awareness of LLMs degrades their performance on RAG tasks. Existing methods to enhance context awareness are often inefficient, incurring time or memory overhead during inference, and many are tailored to specific position embeddings. In this paper, we propose \textbf{P}osition-\textbf{E}mbedding-\textbf{A}gnostic attention \textbf{R}e-weighting (\textit{PEAR}), which enhances the context awareness of LLMs with zero inference overhead. Specifically, on a proxy task focused on context copying, we first detect heads which suppress the models' context awareness, thereby diminishing RAG performance. To weaken the impact of these heads, we re-weight their outputs with learnable coefficients. The LLM (with frozen parameters) is optimized by adjusting these coefficients to minimize loss on the proxy task. During inference, the optimized coefficients are fixed to re-weight these heads, regardless of the specific task at hand. Our proposed \textit{PEAR} offers two major advantages over previous approaches: (1) It introduces zero additional inference overhead in terms of memory usage or inference time, while outperforming competitive baselines in accuracy and efficiency across various RAG tasks. (2) It is independent of position embedding algorithms, ensuring broader applicability.

Deep recommender systems rely heavily on large embedding tables to handle high-cardinality categorical features such as user/item identifiers, and face significant memory constraints at scale. To tackle this challenge, hashing techniques are often employed to map multiple entities to the same embedding and thus reduce the size of the embedding tables. Concurrently, graph-based collaborative signals have emerged as powerful tools in recommender systems, yet their potential for optimizing embedding table reduction remains unexplored. This paper introduces GraphHash, the first graph-based approach that leverages modularity-based bipartite graph clustering on user-item interaction graphs to reduce embedding table sizes. We demonstrate that the modularity objective has a theoretical connection to message-passing, which provides a foundation for our method. By employing fast clustering algorithms, GraphHash serves as a computationally efficient proxy for message-passing during preprocessing and a plug-and-play graph-based alternative to traditional ID hashing. Extensive experiments show that GraphHash substantially outperforms diverse hashing baselines on both retrieval and click-through-rate prediction tasks. In particular, GraphHash achieves on average a 101.52% improvement in recall when reducing the embedding table size by more than 75%, highlighting the value of graph-based collaborative information for model reduction.

As people increasingly use social media as a primary news source, it becomes critical to understand how online platforms affect peoples' experience of the news. Through the media effects of agenda-setting and framing, different news sources can vary in their influence on public opinion regarding which issues people consider important and how particular aspects of these issues should be interpreted. However, little is known about how issues and frames shift and segregate across partisan lines as traditional news on social media gets filtered by the selective exposure effects of social media. In this study, we investigate the issues and frames invoked in news article shares across Reddit over 16 years and measure their traditional media and social media partisanship. We measure the change between production (news articles posted on Reddit) and consumption (news articles posted on Reddit, weighted by their score). We find that issues are shared in a co-partisan manner across traditional media and social media lines. Issues are also more polarized in social media than traditional media and more polarized in consumption than production. We find that frames across several issues are also subject to co-partisan sharing behavior. In contrast to the significant polarization of news outlets on Reddit in 2016, issues and frames do not polarize more over time. Finally, looking at case studies of frames within specific issues, we disaggregate the shift from production to consumption by distinguishing between issues where the frames polarize and issues that simply receive less exposure on one side of the political spectrum. Our results give insight into broader phenomena like political polarization by highlighting the dimensions of precisely what polarizes and how polarization occurs. Overall, our study showcases the importance of understanding how social media distorts the perception of the news via its agenda-setting and framing functions.

Online ads are a source of revenue for millions of websites. However, their intrusive and disruptive nature can impact the user experience of site visitors. Specialized tools such as browser extensions have emerged that block such advertisements from displaying. To restore balance in the favor of domain owners who lost revenue due to ad-filtering, online ad standards were defined to strike a middle ground between user choice and monetization. This paper presents a comprehensive analysis of the compliance of online digital advertisements with the most prevailing ad standard: the Acceptable Ads Standards. We selected 10,000 domains by intersecting Tranco's top 100K domains with the Acceptable Ads exception list. This subset highlights popular sites that are expected to adhere to specific advertising standards. The Acceptable Ads Standards, initiated by the Acceptable Ads Committee, seeks a balance between user experience and ad effectiveness, allowing certain non-intrusive ads defined by size, placement and type limitations. Our research methodology includes a quantitative analysis of ad formats and compliance rates. In this study, we conclude that almost 10\% of the partner websites when crawled with Acceptable Ads' exception list have at-least one non-compliant ad on the landing page. Our analysis also reveals the design flaws in Acceptable Ads Exception list that allows publishers to bypass ad size and format limitations. Leveraging this understanding, we also propose improvements to the exception list that can avoid violating ads from being rendered and ensure user experience of millions of site visitors who rely on Acceptable Ads is improved.

Time series classification (TSC) is a cornerstone of modern web applications, powering tasks such as financial data analysis, network traffic monitoring, and user behavior analysis. In recent years, deep neural networks (DNNs) have greatly enhanced the performance of TSC models in these critical domains. However, DNNs are vulnerable to backdoor attacks, where attackers can covertly implant triggers into models to induce malicious outcomes. Existing backdoor attacks targeting DNN-based TSC models remain elementary. In particular, early methods borrow trigger designs from computer vision, which are ineffective for time series data. More recent approaches utilize generative models for trigger generation, but at the cost of significant computational complexity. In this work, we analyze the limitations of existing attacks and introduce an enhanced method, *FreqBack*. Drawing inspiration from the fact that DNN models inherently capture frequency domain features in time series data, we identify that improper perturbations in the frequency domain are the root cause of ineffective attacks. To address this, we propose to generate triggers both effectively and efficiently, guided by frequency analysis. FreqBack exhibits substantial performance across five models and eight datasets, achieving an impressive attack success rate of over 90%, while maintaining less than a 3% drop in model accuracy on clean data.

Phishing attacks continue to be a major threat to internet users, causing data breaches, financial losses, and identity theft. This study provides an in-depth analysis of the lifespan and evolution of phishing websites, focusing on their survival strategies and evasion techniques. We analyze 286,237 unique phishing URLs over five months using a custom web crawler based on Puppeteer and Chromium. Our crawler runs on a 30-minute cycle, systematically checking the operational status of phishing websites by collecting their HTTP status codes, screenshots, HTML, and HTTP data. Temporal and survival analyses, along with statistical tests, are used to examine phishing website lifecycles, evolution, and evasion tactics. Our findings show that the average lifespan of phishing websites is 54 hours (2.25 days) with a median of 5.46 hours, indicating rapid takedown of many sites while a subset remains active longer. Interestingly, logistic-themed phishing websites (e.g., USPS) operate within a compressed timeframe (1.76 hours) compared to other brands (e.g., Facebook). We further analyze detection effectiveness using Google Safe Browsing (GSB). We find that GSB detects only 18.4% of phishing websites, taking an average of 4.5 days. Notably, 83.93% of phishing sites are already taken down before GSB detection, meaning GSB requires more prompt detection. Moreover, 16.07% of phishing sites persist beyond this point, surviving for an additional 7.2 days on average, resulting in an average total lifespan of approximately 12 days. We reveal that DNS resolution error is the main cause (67%) of phishing website takedowns. Finally, we uncover that phishing sites with extensive visual changes (more than 100 times) exhibit a median lifespan of 17 days, compared to 1.93 hours for those with minimal modifications. These results highlight the dynamic nature of phishing attacks, the challenges in detection and prevention, and the need for more rapid and comprehensive countermeasures against evolving phishing tactics.

Recent advancements have highlighted that Large Language Models (LLMs) are prone to hallucinations when solving complex reasoning problems, leading to erroneous results. To tackle this issue, researchers incorporate Knowledge Graphs (KGs) to improve the reasoning ability of LLMs. However, existing methods face two limitations: 1) they typically assume that all answers to the questions are contained in KGs, neglecting the incompleteness issue of KGs, and 2) they treat the KG as a static repository and overlook the implicit logical reasoning structures inherent in KGs. In this paper, we introduce SymAgent, an innovative neural-symbolic agent framework that achieves collaborative augmentation between KGs and LLMs. We conceptualize KGs as dynamic environments and transform complex reasoning tasks into a multi-step interactive process, enabling KGs to participate deeply in the reasoning process. SymAgent consists of two modules: Agent-Planner and Agent-Executor. The Agent-Planner leverages LLM's inductive reasoning capability to extract symbolic rules from KGs, guiding efficient question decomposition. The Agent-Executor autonomously invokes predefined action tools to integrate information from KGs and external documents, addressing the issues of KG incompleteness. Furthermore, we design a self-learning framework comprising online exploration and offline iterative policy updating phases, enabling the agent to automatically synthesize reasoning trajectories and improve performance. Experimental results demonstrate that SymAgent with weak LLM backbones (i.e., 7B series) yields better or comparable performance compared to various strong baselines. Further analysis reveals that our agent can identify missing triples, facilitating automatic KG updates. The code is available at \url{https://anonymous.4open.science/r/SymAgent/}.

Spatial data play an important role in many applications built over knowledge graphs, and are frequently referenced in queries posed to public query services, such as that of Wikidata. Querying for spatial data presents a significant challenge, as dealing with topological relations such as adjacent or contains implies dealing with inferred information, such as through the transitivity of the containment relation. However, despite all the recent advances in querying knowledge graphs, we still lack techniques specifically tailored for topological information. Applications looking to incorporate topological relations must either materialize the inferred relations, incurring high space and maintenance overheads, or query them with less efficient recursive algorithms, incurring high runtime overheads. In this paper we address the problem of leveraging topological information in knowledge graphs by designing efficient algorithms to process these queries. Our solution involves building a specific index that stores the topological information in a convenient compact form, and includes specialized algorithms that infer every possible relation from the basic topological facts in the graph. We show that, while using essentially the same space required to solve standard graph pattern queries, we can incorporate topological predicates, accounting for all the inferred information, all within worst-case-optimal time. We implement our scheme and show experimentally that it outperforms baseline solutions by a notable margin.

Dynamic graph clustering aims to detect and track time-varying clusters in dynamic graphs, revealing the evolutionary mechanisms of complex real-world dynamic systems. Matrix factorization-based methods are promising approaches for this task; however, these methods often struggle with scalability and can be time-consuming when applied to large-scale dynamic graphs. Moreover, they tend to lack robustness and are vulnerable to real-world noisy data. To address these issues, we make three key contributions. First, to improve scalability, we propose temporal separated matrix factorization, where a single matrix is divided into multiple smaller matrices for independent factorization, resulting in faster computation. Second, to improve robustness, we introduce bi-clustering regularization, which jointly optimizes graph embedding and clustering, thereby filtering out noisy features from the graph embeddings. Third, to further enhance effectiveness and efficiency, we propose selective embedding updating, where we update only the embeddings of dynamic nodes while the embeddings of static nodes are fixed among different timestamps. Experimental results on six synthetic and five real-world benchmarks demonstrate the scalability, robustness and effectiveness of our proposed method. Our code is available on Github https://anonymous.4open.science/r/RS-MF-50FE/README.md.

Dynamic graph neural networks (DGNNs) have emerged and been widely deployed in various web applications (e.g., Reddit) to serve users (e.g., personalized content delivery) due to their remarkable ability to learn from complex and dynamic user interaction data. Despite benefiting from high-quality services, users have raised privacy concerns, such as misuse of personal data (e.g., dynamic user-user/item interaction) for model training, requiring DGNNs to ''forget'' their data to meet AI governance laws (e.g., the ''right to be forgotten" in GDPR). However, current static graph unlearning studies cannot $\textit{unlearn dynamic graph elements}$ and exhibit limitations such as the model-specific design or reliance on pre-processing, which disenable their practicability in dynamic graph unlearning. To this end, we study the dynamic graph unlearning for the first time and propose an $\textit{effective}$, $\textit{efficient}$, $\textit{general}$, and $\textit{post-processing}$ method to implement DGNN unlearning. Specifically, we first formulate dynamic graph unlearning in the context of continuous-time dynamic graphs, and then propose a method called Gradient Transformation that directly maps the unlearning request to the desired parameter update. Comprehensive evaluations on six real-world datasets and state-of-the-art DGNN backbones demonstrate its effectiveness (e.g., limited drop or obvious improvement in utility) and efficiency (e.g., 7.23$\times$ speed-up) advantages. Additionally, our method has the potential to handle future unlearning requests with significant performance gains (e.g., 32.59$\times$ speed-up).

Phishing attacks on Ethereum have increased with its growing adoption, creating significant challenges as phishing and non-phishing users often display similar behavior. Additionally, while the network as a whole experiences high activity, individual user behavior is typically sparse, making it difficult to detect phishing patterns. Current methods frequently fail to tackle these challenges and often neglect the temporal sequence of transactions, resulting in data leakage and reduced performance. In this paper, we propose a novel approach that addresses these gaps by focusing on the association of two key aspects: (1) local temporal behavior fluctuations of individual users and (2) deviations from global transaction patterns within the network. To aim this, we introduce CATALOG (CApturing joint TemporAl dependencies from LOcal and Global user behaviour), a novel representation learning model that jointly captures the local and global behavioral patterns of a user and their correlations by leveraging a dual cross-attention mechanism paired with a bi-directional Masked Language Modelling (MLM) based pipelined transformer framework. Our proposed model simultaneously learns from local behavioral shifts and global market trends along with a contextually enriched embeddings, effectively distinguishing phishing from non-phishing users, while addressing the existing research gaps. Extensive experiments on real-world Ethereum transaction data show that our framework improves phishing detection by 7-8% in F1-Score compared to existing models. Furthermore, it generalizes effectively across Ethereum versions 1.0 and 2.0, demonstrating the robustness of our approach.

Governments and regulatory bodies have recognized investment scams as the most prevalent forms of cryptocurrency fraud. These scams typically use professional-looking websites to lure unsuspecting victims with promises of unrealistically high returns. In this paper, we introduce Crimson, a distributed system designed to continuously detect cryptocurrency investment scam websites as they are created in the wild. Over the first 8 months of 2024, Crimson processed approximately 6 billion domain names and classified 43, 572 unique cryptocurrency investment scam websites in real-time. Beyond detection, we provide insights into the design and infrastructure of these websites that can help users recognize scam patterns and assist hosting providers in detecting and blocking such sites. Among others, we discovered that most investment scam websites use similar templates and that 52% of all scam websites were hosted on just 10% of all resolved IP addresses, indicating a concentration of scam operations within a small subset of hosting providers. Furthermore, we investigate the inclusion of our detected scam websites in blacklists used by popular web browsers and applications, finding that the vast majority of these websites were absent. On the financial side, by analyzing the incoming transactions to scammer wallets on 6.7% of the sites detected by Crimson, we observe an estimated lower bound of 2.04M USD in losses because of cryptocurrency investment scams, pointing to tens of millions of dollars of losses in total.

In this paper, we perform a systematic study to explore a pivotal problem facing the web community: is current distributed web cache ready for future satellite Internet? First, through a worldwide performance measurement based on the RIPE Atlas platform and Starlink, the largest low-earth orbit (LEO) satellite network (LSN) today, we identify that the uneven deployment of current distributed cache servers, inter-ISP meandering routes and the last-mile congestion on LEO links prevent existing terrestrial web cache from providing low-latency web access for users in emerging LSNs. Second, we propose Spache, a novel web caching system which addresses the limitations of existing ground-only cache by exploiting a bold idea: integrating web cache into LEO satellites to achieve ubiquitous and low-latency web services. Specifically, Spache leverages a key feature of LSNs called communication schedule to efficiently prefetch web contents on satellites, and adopts a schedule-driven partitioning strategy to avoid cache pollution involved by LEO mobility. Finally, we implement a prototype of Spache, and evaluate it based on real-world HTTP traces and real-data-driven LSN simulation. Extensive evaluations demonstrate that as compared to existing distributed caching solutions, Spache can improve cache hit ratio by 19.8% on average, reduce latency by up to 17.7%, and sustain consistently low user-to-cache latency for global LSN users.

Representation learning technique is an effective link prediction paradigm to alleviate the incompleteness of knowledge hypergraphs. However, the $n$-ary complex semantic information inherent in knowledge hypergraphs causes existing methods to face the dual limitations of weak effectiveness and low efficiency. In this paper, we propose a novel knowledge hypergraph representation learning model, HySAE, which can achieve a satisfactory trade-off between effectiveness and efficiency. Concretely, HySAE builds an efficient semantic-enhanced 3D scalable end-to-end embedding architecture to sufficiently capture knowledge hypergraph $n$-ary complex semantic information with fewer parameters, which can significantly reduce the computational cost of the model. In particular, we also design an efficient position-aware entity role semantic embedding way and two enhanced semantic learning strategies to further improve the effectiveness and scalability of our proposed method. Extensive experimental results on all datasets demonstrate that HySAE consistently outperforms state-of-the-art baselines, with an average improvement of 9.15\%, a maximum improvement of 39.44\%, an average 10.39x faster, and 75.79\% fewer parameters.

With the proliferation of images in online content, language-guided image retrieval (LGIR) has emerged as a research hotspot over the past decade, encompassing a variety of subtasks with diverse input forms. While the development of large multimodal models (LMMs) has significantly facilitated these tasks, existing approaches often address them in isolation, requiring the construction of separate systems for each task. This not only increases system complexity and maintenance costs, but also exacerbates challenges stemming from language ambiguity and complex image content, making it difficult for retrieval systems to provide accurate and reliable results. To this end, we propose ImageScope, a training-free, three-stage framework that leverages collective reasoning to unify LGIR tasks. The key insight behind the unification lies in the compositional nature of language, which transforms diverse LGIR tasks into a generalized text-to-image retrieval process, along with the reasoning of LMMs serving as a universal verification to refine the results. To be specific, in the first stage, we improve the robustness of the framework by synthesizing search intents across varying levels of semantic granularity using chain-of-thought (CoT) reasoning. In the second and third stages, we then reflect on retrieval results by verifying predicate propositions locally, and performing pairwise evaluations globally. Experiments conducted on six LGIR datasets demonstrate that ImageScope outperforms competitive baselines. Comprehensive evaluations and ablation studies further confirm the effectiveness of our design.

In today's technology-driven world, web services have opened up new opportunities for blind and visually impaired people to interact independently. Securing interactions with these services is crucial; however, currently deployed methods of web authentication mainly concentrate on sighted users, overlooking the specific needs of the blind and visually impaired community. In this paper, we address this critical gap by investigating the security and accessibility aspects of these web authentication methods when adopted by blind and visually impaired users. We model web authentication for such users as screen reader assisted authentication and introduce an evaluation framework called Authentication Workflows Accessibility Review and Evaluation (AWARE). Using AWARE, we then systematically assessed popular PC-based and smartphone-based screen readers against different types of deployed web authentication methods, including variants of 2FA and passwordless schemes, to simulate real-world scenarios for blind and visually impaired individuals. We analyzed these screen reader assisted authentication interactions with authentication methods in three settings: using a terminal (PC) with screen readers, a combination of the terminal (PC) and smartphone with screen readers, and smartphones with integrated screen readers. The results of our study underscore significant weaknesses in all of our observed screen reader assisted authentication scenarios for real-life authentication methods. These weaknesses, encompassing specific accessibility issues caused by imprecise screen reader instructions, highlight vulnerability concerning observed scenarios for both real-world and research literature based attacks, including phishing, concurrency, fatigue, cross-service, and shoulder surfing. Broadly, our AWARE framework can be used by authentication system designers as a precursor to user studies which are typically time-consuming and tedious to perform, independently allowing to unfold security and accessibility problems early which designers can address prior to full-fledged user testing of more isolated issues.