Multilingual Visual Document Retrieval (VDR) aims to retrieve relevant document screenshots given a textual query across different languages. Recent late-interaction VDR models estimate relevance via token-level maximum similarity aggregation (MaxSim), but this heuristic can introduce semantic bias in multilingual settings. In particular, linguistic differences such as morphological variation, functional word distribution, and tokenization granularity can lead to heterogeneous query token representations. When aggregated under MaxSim, such heterogeneity may result in disproportionate token contributions and inconsistent relevance estimation across languages. To address these limitations, we propose Language-Vision-Aware Late Interaction (LVALI), a lightweight calibration framework without modifying the original embeddings. LVALI introduces two complementary mechanisms: (1) a language-aware token weighting scheme that calibrates token-level contributions at the vocabulary level, and (2) a visual alignment stability-based correction term that downweights low-confidence token-patch matches according to the dispersion of similarity distributions in the document image space. This design preserves the underlying similarity structure of MaxSim while improving robustness to cross-lingual variability. Experiments on multiple multilingual VDR benchmarks demonstrate consistent improvements in robustness and cross-lingual consistency, achieved with only lightweight token-level calibration training.

Multimodal large language models (MLLMs) demonstrate strong vision–language understanding, yet their safety enforcement often lags behind their reasoning capability. Existing multimodal jailbreak methods, such as HADES and FigStep, primarily rely on visual obfuscation or OCR-based injection, and often show limited robustness across models and safety policies. We identify a deeper vulnerability, termed structure-induced safety gaps, in which models retain semantic comprehension but safety enforcement weakens when generation is guided by structured layouts. By embedding harmful cues into tabular templates and framing tasks as benign completion operations, structured reasoning can weaken refusal triggering and enable harmful outputs even when the underlying content is recognized. Building on this insight, we propose SISG, a layout-driven jailbreak framework that exploits structured multimodal reasoning. Experiments on two multimodal safety benchmarks and six advanced MLLMs, including GPT-4o, show that SISG consistently outperforms strong baselines in attack success while maintaining semantic plausibility. Further behavioral analysis reveals reduced refusal responses and elevated harmful scores in failed cases, indicating systematic weaknesses in current safety pipelines. These findings highlight the need for structure-aware safety alignment in multimodal reasoning systems.

Detecting social media bots is essential for maintaining the security and trustworthiness of social networks. While contemporary graph-based detection methods demonstrate promising results, their practical application is limited by label reliance and poor generalization capability across diverse communities. Generative Graph Self-Supervised Learning (GSL) presents a promising paradigm to overcome these limitations, yet existing approaches predominantly follow the homophily assumption and fail to capture the global patterns in the graph, which potentially diminishes their effectiveness when facing the challenges of interaction camouflage and distributed deployment in bot detection scenarios. To this end, we propose BotHP, a generative GSL framework tailored to boost graph-based bot detectors through heterophily-aware representation learning and prototype-guided cluster discovery. Specifically, BotHP leverages a dual-encoder architecture, consisting of a graph-aware encoder to capture node commonality and a graph-agnostic encoder to preserve node uniqueness. This enables the simultaneous modeling of both homophily and heterophily, effectively countering the interaction camouflage issue. Additionally, BotHP incorporates a prototype-guided cluster discovery pretext task to model the latent global consistency of bot clusters and identify spatially dispersed yet semantically aligned bot collectives. Extensive experiments on two real-world bot detection benchmarks demonstrate that BotHP consistently boosts graph-based bot detectors, improving detection performance, alleviating label reliance, and enhancing generalization capability.

Multi-view clustering has garnered increasing research attention due to its capacity to learn common semantics across views for enhanced performance. Numerous methods have been proposed to effectively learn multi-view common semantics, among which deep learning-based approaches have gradually become mainstream owing to their superior representation capabilities. However, most existing methods only fuse multi-view information at the final clustering stage, causing underutilization of multi-view data during training. To address this limitation, we propose a novel framework that treats fused multi-view features as an additional view during model training (FADE). Our autoencoder-based method first integrates deep features from multiple views to generate new representations, which are incorporated as the (V+1)-th view. Through dedicated MLPs, we extract high-level features and cluster assignments from all views. To learn common semantics, contrastive learning is first conducted between the (V+1)-th view and other views, followed by comprehensive cross-view contrastive learning to enforce multi-view consistency. Finally, we fine-tune the model by jointly optimizing high-level semantic features and cluster assignments. Comparative experiments demonstrate that our method effectively extracts common information while mitigating the impact of view-private information, outperforming multiple approaches.

Recently, federated multi-view clustering has gained attention as an effective approach for exploring the clustering structure of multi-view/multi-modal data distributed across multiple clients. However, most existing methods primarily focus on sample-related client scenarios with limited research on sample-unrelated settings. These scenarios face severe data heterogeneity, while missing data further reduces available knowledge, making the problem more challenging. To address these challenges, we propose FedHMAI, a novel horizontal federated incomplete multi-view clustering method. Specifically, on the server side, we perform heterogeneity matching by leveraging Maximum Mean Discrepancy (MMD) to identify and distribute to each client the features with the least heterogeneity. Then on the client side, local cluster structure learning is conducted to alleviate data heterogeneity and reinforce cluster discriminability. Furthermore, we design an attention-based dual-level imputation mechanism that jointly leverages data- and feature-level imputation, along with matching features, to ensure imputation accuracy and consistency. Extensive experimental results validate that FedHMAI effectively overcomes the challenges of incomplete multi-view data in horizontal federated settings, consistently achieving superior clustering performance.

Under stringent privacy constraints, whether federated recommendation systems can achieve group fairness remains an inadequately explored question. Taking gender fairness as a representative issue, we identify three phenomena in federated recommendation systems: performance difference, data imbalance, and preference disparity. We discover that the state-of-the-art methods only focus on the first phenomenon. Consequently, their imposition of inappropriate fairness constraints detrimentally affects the model training. Moreover, due to insufficient sensitive attribute protection of existing works, we can infer the gender of all users with 99.90% accuracy even with the addition of maximal noise. In this work, we propose Privacy-Preserving Orthogonal Aggregation (PPOA), which employs the secure aggregation scheme and quantization technique, to prevent the suppression of minority groups by the majority and preserve the distinct preferences for better group fairness. PPOA can assist different groups in obtaining their respective model aggregation results through a designed orthogonal mapping while keeping their attributes private. Experimental results on three real-world datasets demonstrate that PPOA enhances recommendation effectiveness for both females and males by up to 8.25% and 6.36%, respectively, with a maximum overall improvement of 7.30%, and achieves optimal fairness in most cases. Extensive ablation experiments and visualizations indicate that PPOA successfully maintains preferences for different gender groups.

Recent advancements in social bot detection have been driven by the adoption of Graph Neural Networks. The social graph, constructed from social network interactions, contains benign and bot accounts that influence each other. However, previous graph-based detection methods that follow the transductive message-passing paradigm may not fully utilize hidden graph information and are vulnerable to adversarial bot behavior. The indiscriminate message passing between nodes from different categories and communities results in excessively homogeneous node representations, ultimately reducing the effectiveness of social bot detectors. In this paper, we propose \SEBot, a novel multi-view graph-based contrastive learning-enabled social bot detector. In particular, we use structural entropy as an uncertainty metric to optimize the entire graph's structure and subgraph-level granularity, revealing the implicitly existing hierarchical community structure. And we design an encoder to enable message passing beyond the homophily assumption, enhancing robustness to adversarial behaviors of social bots. Finally, we employ multi-view contrastive learning to maximize mutual information between different views and enhance the detection performance through multi-task learning. Experimental results demonstrate that our approach significantly improves the performance of social bot detection compared with SOTA methods.

Privacy protection in recommendation systems is gaining increasing attention, for which federated learning has emerged as a promising solution. Current federated recommendation systems grapple with high communication overhead due to sharing dense global embeddings, and also poorly reflect user preferences due to data heterogeneity. To overcome these challenges, we propose a two-stage Federated Low-rank Coordinated Adaptation (FedLoCA) framework to decouple global and client-specific knowledge into low-rank embeddings, which significantly reduces communication overhead while enhancing the system’s ability to capture individual user preferences amidst data heterogeneity. Further, to tackle gradient estimation inaccuracies stemming from data sparsity in federated recommendation systems, we introduce an adversarial gradient projected descent approach in low-rank spaces, which significantly boosts model performance while maintaining robustness. Remarkably, FedLoCA also alleviates performance loss even under the stringent constraints of differential privacy. Extensive experiments on various real-world datasets demonstrate that FedLoCA significantly outperforms existing methods in both recommendation accuracy and communication efficiency.

Social bot detection is of paramount importance to the resilience and security of online social platforms. The state-of-the-art detection models are siloed and have largely overlooked a variety of data characteristics from multiple cross-lingual platforms. Meanwhile, the heterogeneity of data distribution and model architecture make it intricate to devise an efficient cross-platform and cross-model detection framework. In this paper, we propose FedACK, a new federated adversarial contrastive knowledge distillation framework for social bot detection. We devise a GAN-based federated knowledge distillation mechanism for efficiently transferring knowledge of data distribution among clients. In particular, a global generator is used to extract the knowledge of global data distribution and distill it into each client’s local model. We leverage local discriminator to enable customized model design and use local generator for data enhancement with hard-to-decide samples. Local training is conducted as multi-stage adversarial and contrastive learning to enable consistent feature spaces among clients and to constrain the optimization direction of local models, reducing the divergences between local and global models. Experiments demonstrate that FedACK outperforms the state-of-the-art approaches in terms of accuracy, communication efficiency, and feature space consistency.

Recommender systems should answer the intervention question "if recommending an item to a user, what would the feedback be", calling for estimating the causal effect of a recommendation on user feedback. Generally, this requires blocking the effect of confounders that simultaneously affect the recommendation and feedback. To mitigate the confounding bias, a strategy is incorporating propensity into model learning. However, existing methods forgo possible unmeasured confounders (e.g., user financial status), which can result in biased propensities and hurt recommendation performance. This work combats the risk of unmeasured confounders in recommender systems.

Most recommender systems evaluate model performance offline through either: 1) normal biased test on factual interactions; or 2) debiased test with records from the randomized controlled trial. In fact, both tests only reflect part of the whole picture: factual interactions are collected from the recommendation policy, fitting them better implies benefiting the platform with higher click or conversion rate; in contrast, debiased test eliminates system-induced biases and thus is more reflective of user true preference. Nevertheless, we find that existing models exhibit trade-off on the two tests, and there lacks methods that perform well on both tests.

The fast pace of global urbanization is drastically changing the population distributions over the world, which leads to significant changes in geographical population densities. Such changes in turn alter the underlying geographical power demand over time, and drive power substations to become over-supplied (demand << capacity) or under-supplied (demand ≈ capacity). In this paper, we make the first attempt to investigate the problem of power substation-user assignment by analyzing large-scale power grid data. We develop a Scalable Power User Assignment (SPUA) framework, that takes large-scale spatial power user/substation distribution data and temporal user power consumption data as input, and assigns users to substations, in a manner that minimizes the maximum substation utilization among all substations. To evaluate the performance of our SPUA framework, we conduct evaluations on real power consumption data and user/substation location data collected from a province in China for 35 days in 2015. The evaluation results demonstrate that our SPUA framework can achieve a 20%--65% reduction on the maximum substation utilization, and 2 to 3.7 times reduction on total transmission loss over other baseline methods.