In-context learning (ICL) has proven to be adept at adapting large language models (LLMs) to downstream tasks without parameter updates, based on a few demonstration examples. Prior work has found that the ICL performance is susceptible to the selection of examples in prompt and made efforts to stabilize it. However, existing example selection studies ignore the ethical risks behind the examples selected, such as gender and race bias. In this work, we conduct extensive experiments and discover that (1) example selection with high accuracy does not mean low bias; (2) example selection for ICL may amplify the biases of LLMs; (3) example selection contributes to spurious correlations of LLMs. Based on the above observations, we propose the Remind with Bias-aware Embedding (ReBE), which removes the spurious correlations through contrastive learning and obtains bias-aware embedding for LLMs based on prompt tuning. Finally, we demonstrate that ReBE effectively mitigates biases of LLMs without significantly compromising accuracy and is highly compatible with existing example selection methods.

With the rapid deployment of Chinese large language models (LLMs), culturally-grounded bias evaluation remains understudied due to the dominance of English benchmarks and simplistic Chinese scenarios. To address this, we propose GeWu, a comprehensive benchmark featuring a culturally-aware dataset of 60,192 questions spanning 14 social groups with fine-grained Chinese contexts, significantly exceeding existing resources in breadth and depth. Our two-stage evaluation first quantifies bias via multiple-choice questions using a novel probability-based scoring mechanism to sensitively capture bias tendencies, distilling high-bias scenarios into GeWu-1K. This refined subset then enables multi-turn dialogue evaluations for in-depth analysis under realistic conditions. Experiments reveal that GeWu effectively exposes social biases in state-of-the-art Chinese LLMs, with 13.93% of scenarios eliciting universal bias across all models. This highlights persistent challenges and provides actionable insights for bias mitigation in Chinese contexts.

Tensor network structure search (TN-SS) aims to automatically discover optimal network topologies and rank configurations for efficient tensor decomposition in high-dimensional data representation. Despite recent advances, existing TN-SS methods face significant limitations in computational tractability, structure adaptivity, and optimization robustness across diverse tensor characteristics. They struggle with three key challenges: single-scale optimization missing multi-scale structures, discrete search spaces hindering smooth structure evolution, and separated structure-parameter optimization causing computational inefficiency. We propose RGTN (Renormalization Group guided Tensor Network search), a physics-inspired framework transforming TN-SS via multi-scale renormalization group flows. Unlike fixed-scale discrete search methods, RGTN uses dynamic scale-transformation for continuous structure evolution across resolutions. Its core innovation includes learnable edge gates for optimization-stage topology modification and intelligent proposals based on physical quantities like node tension measuring local stress, and edge information flow quantifying connectivity importance. Starting from low-complexity coarse scales and refining to finer ones, RGTN finds compact structures while escaping local minima via scale-induced perturbations. Extensive experiments on light field data, high-order synthetic tensors, and video completion tasks show RGTN achieves state-of-the-art compression ratios and runs 4-600 times faster than existing methods, validating the effectiveness of our physics-inspired approach.

Fine-grained urban flow inference is crucial for urban planning and intelligent transportation systems, enabling precise traffic management and resource allocation. However, the practical deployment of existing methods is hindered by two key challenges: the prohibitive computational cost of over-parameterized models and the suboptimal performance of conventional loss functions on the highly skewed distribution of urban flows. To address these challenges, we propose a unified solution that synergizes architectural efficiency with adaptive optimization. Specifically, we first introduce PLGF, a lightweight yet powerful architecture that employs a Progressive Local-Global Fusion strategy to effectively capture both fine-grained details and global contextual dependencies. Second, we propose DualFocal Loss, a novel function that integrates dual-space supervision with a difficulty-aware focusing mechanism, enabling the model to adaptively concentrate on hard-to-predict regions. Extensive experiments on 4 real-world scenarios validate the effectiveness and scalability of our method. Notably, while achieving state-of-the-art performance, PLGF reduces the model size by up to 97% compared to current high-performing methods. Furthermore, under comparable parameter budgets, our model yields an accuracy improvement of over 10% against strong baselines.

Xiaopeng Li, Yuanjin Zheng, Wanyu Wang, Wenlin Zhang, Pengyue Jia, Yingyi Zhang, Haiying He, Mengyang Ma, Yiqi Wang, Maolin Wang, Xuetao Wei, Xiangyu Zhao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Machine unlearning for large language models (LLMs) aims to remove targeted knowledge while preserving general capability.In this paper, we recast LLM unlearning as an asymmetric two-task problem: retention is the primary objective and forgetting is an auxiliary.From this perspective, we propose a retention-prioritized gradient synthesis framework that decouples task-specific gradient extraction from conflict-aware combination.Instantiating the framework, we adapt established PCGrad to resolve gradient conflicts, and introduce SAGO, a novel retention-prioritized gradient synthesis method.Theoretically, both variants ensure non-negative cosine similarity with the retain gradient, while SAGO achieves strictly tighter alignment through constructive sign-constrained synthesis.Empirically, on WMDP Bio/Cyber and RWKU benchmarks, SAGO consistently pushes the Pareto frontier: e.g., on WMDP Bio (SimNPO+GD), recovery of target model MMLU performance progresses from 44.6% (naive) to 94.0% (+PC-Grad) and further to 96.0% (+SAGO), while maintaining comparable forgetting strength.Our results show that re-shaping gradient geometry, rather than re-balancing losses, is the key to mitigating unlearning-retention trade-offs.

Large vision-language models (LVLMs) have shown significant progress in video understanding, but they struggle to scale to long videos due to limited context windows. Existing methods reduce input dimensionality via frame sampling and feature compression, yet discard details and incur high computational cost for post-training. In contrast, retrieval-augmented generation (RAG) that indexes long videos for query retrieval and memory-based methods that maintain evolving long-term stores, offer a lighter and deployment-friendly solution. Nevertheless, they rely on shallow retrieval that selects only top-ranked segments and fails to integrate information across multiple relevant video episodes. Inspired by Multiple-Trace Theory in cognitive psychology, we revisit long video understanding from a probe-echo perspective, in which human episodic memories are activated and integrated in parallel. Building on this insight, we propose ProEchoMem, a cognitive-inspired framework that simulates the probe-echo mechanism: (1) Incremental Episodic Memory Construction builds structured knowledge graphs from video streams; (2) Probe-Driven Memory Activation generates probe signals from user queries to activate all stored traces simultaneously; (3) Memory Echo Synthesis integrates activated traces into a coherent and structured memory echo. Experiments on LongerVideos, LVBench, and cross-domain settings demonstrate the effectiveness of ProEchoMem, with multi-trace probing achieving up to 14.2% higher relevance and ablation studies validating the contribution of each module. The code is available at https://github.com/Applied-Machine-Learning-Lab/SIGIR26_ProEchoMem

As urban residents demand higher travel quality, vehicle dispatch has become a critical component of online ride-hailing services. However, current vehicle dispatch systems struggle to navigate the complexities of urban traffic dynamics, including unpredictable traffic conditions, diverse driver behaviors, and fluctuating supply and demand patterns. These challenges have resulted in travel difficulties for passengers in certain areas, while many drivers in other areas are unable to secure orders, leading to a decline in the overall quality of urban transportation services. To address these issues, this paper introduces GARLIC: a framework of GPT-Augmented Reinforcement Learning with Intelligent Control for vehicle dispatching. GARLIC utilizes multiview graphs to capture hierarchical traffic states, and learns a dynamic reward function that accounts for individual driving behaviors. The framework further integrates a GPT model trained with a custom loss function to enable high-precision predictions and optimize dispatching policies in real-world scenarios. Experiments conducted on two real-world datasets demonstrate that GARLIC effectively aligns with driver behaviors while reducing the empty load rate of vehicles.

Recent advancements in text-to-3D generation can generate neural radiance fields (NeRFs) with score distillation sampling, enabling 3D asset creation without real-world data capture. With the rapid advancement in NeRF generation quality, protecting the copyright of the generated NeRF has become increasingly important. While prior works can watermark NeRFs in a post-generation way, they suffer from two vulnerabilities. First, a delay lies between NeRF generation and watermarking because the secret message is embedded into the NeRF model post-generation through fine-tuning. Second, generating a non-watermarked NeRF as an intermediate creates a potential vulnerability for theft. To address both issues, we propose Dreamark to embed a secret message by backdooring the NeRF during NeRF generation. In detail, we first pre-train a watermark decoder. Then, Dreamark generates backdoored NeRFs in a way that the target secret message can be verified by the pre-trained watermark decoder on an arbitrary trigger viewport. We evaluate the generation quality and watermark robustness against image- and model-level attacks. Extensive experiments show that the watermarking process will not degrade the generation quality, and the watermark achieves 90+% accuracy among both image-level attacks (e.g., Gaussian noise) and model-level attacks (e.g., pruning attack).

Mesh watermark embeds secret messages in 3D meshes and decodes the message from watermarked meshes for ownership verification. Current watermarking methods directly hide secret messages in vertex and face sets of meshes. However, mesh is a discrete representation that uses vertex and face sets to describe a continuous signal, which can be discretized in other discrete representations with different vertex and face sets. This raises the question of whether the watermark can still be verified on the different discrete representations of the watermarked mesh. We conduct this research in an attack-then-defense manner by proposing a novel function space mesh watermark removal attack FuncEvade and then mitigating it through function space mesh watermarking FuncMark. In detail, FuncEvade generates a different discrete representation of a watermarked mesh by extracting it from the signed distance function of the watermarked mesh. We observe that the generated mesh can evade ALL previous watermarking methods. FuncMark mitigates FuncEvade by watermarking signed distance function through message-guided deformation. Such deformation can survive isosurfacing and thus be inherited by the extracted meshes for further watermark decoding. Extensive experiments demonstrate that FuncEvade achieves 100% evasion rate among all previous watermarking methods while achieving only 0.3% evasion rate on FuncMark. Besides, our FuncMark performs similarly on other metrics compared to state-of-the-art mesh watermarking methods.

Yuxuan Li, Xinwei Guo, Jiashi Gao, Guanhua Chen, Xiangyu Zhao, Jiaxin Zhang, Quanying Liu, Haiyan Wu, Xin Yao, Xuetao Wei. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

With the proliferation of task-specific large language models, delta compression has emerged as a method to mitigate the resource challenges of deploying numerous such models by effectively compressing the delta model parameters. Previous delta-sparsification methods either remove parameters randomly or truncate singular vectors directly after singular value decomposition (SVD). However, these methods either disregard parameter importance entirely or evaluate it with too coarse a granularity. In this work, we introduce ImPart, a novel importance-aware delta sparsification approach. Leveraging SVD, it dynamically adjusts sparsity ratios of different singular vectors based on their importance, effectively retaining crucial task-specific knowledge even at high sparsity ratios. Experiments show that ImPart achieves state-of-the-art delta sparsification performance, demonstrating 2\times higher compression ratio than baselines at the same performance level. When integrated with existing methods, ImPart sets a new state-of-the-art on delta quantization and model merging.

Neural Architecture Search (NAS) has emerged as a powerful approach for automating neural network design. However, existing NAS methods face critical limitations in real-world deployments: architectures lack adaptability across scenarios, each deployment context requires costly separate searches, and performance consistency across diverse platforms remains challenging. We propose DANCE (Dynamic Architectures with Neural Continuous Evolution), which reformulates architecture search as a continuous evolution problem through learning distributions over architectural components. DANCE introduces three key innovations: a continuous architecture distribution enabling smooth adaptation, a unified architecture space with learned selection gates for efficient sampling, and a multi-stage training strategy for effective deployment optimization. Extensive experiments across five datasets demonstrate DANCE's effectiveness. Our method consistently outperforms state-of-the-art NAS approaches in terms of accuracy while significantly reducing search costs. Under varying computational constraints, DANCE maintains robust performance while smoothly adapting architectures to different hardware requirements. The code and appendix can be found at https://github.com/Applied-Machine-Learning-Lab/DANCE.

Online advertising in recommendation platforms has gained significant attention, with a predominant focus on channel recommendation and budget allocation strategies. However, current offline reinforcement learning (RL) methods face substantial challenges when applied to sparse advertising scenarios, primarily due to severe overestimation, distributional shifts, and overlooking budget constraints. To address these issues, we propose MTORL, a novel multi-task offline RL model that targets two key objectives. First, we establish a Markov Decision Process (MDP) framework specific to the nuances of advertising. Then, we develop a causal state encoder to capture dynamic user interests and temporal dependencies, facilitating offline RL through conditional sequence modeling. Causal attention mechanisms are introduced to enhance user sequence representations by identifying correlations among causal states. We employ multi-task learning to decode actions and rewards, simultaneously addressing channel recommendation and budget allocation. Notably, our framework includes an automated system for integrating these tasks into online advertising. Extensive experiments on offline and online environments demonstrate MTORL's superiority over state-of-the-art methods. The code is available online at https://github.com/Applied-Machine-Learning-Lab/MTORL.

Transformer-based models have gained significant traction in sequential recommender systems (SRSs) for their ability to capture user-item interactions effectively. However, these models often suffer from high computational costs and slow inference. Meanwhile, existing efficient SRS approaches struggle to embed high-quality semantic and positional information into latent representations. To tackle these challenges, this paper introduces GLINT-RU, a lightweight and efficient SRS leveraging a single-layer dense selective Gated Recurrent Units (GRU) module to accelerate inference. By incorporating a dense selective gate, GLINT-RU adaptively captures temporal dependencies and fine-grained positional information, generating high-quality latent representations. Additionally, a parallel mixing block infuses fine-grained positional features into user-item interactions, enhancing both recommendation quality and efficiency. Extensive experiments on three datasets demonstrate that GLINT-RU achieves superior prediction accuracy and inference speed, outperforming baselines based on RNNs, Transformers, MLPs, and SSMs. These results establish GLINT-RU as a powerful and efficient solution for SRSs. The implementation code is publicly available for reproducibility. https://github.com/szhang-cityu/GLINT-RU.

Text-to-Image (T2I) diffusion models exhibit concerning tendencies to generate harmful imagery that perpetuates social biases and stereotypes, posing significant ethical risks in real-world applications. While existing mitigation approaches predominantly employ black-box methodologies through dataset augmentation or constrained fine-tuning, they face critical limitations, including high data acquisition costs and potential exacerbation of stereotypes during model retraining. Inspired by neuroscience principles where neurological dysfunction often stems from aberrant neural activation patterns, we propose a novel framework, StereoClinic, targeting the root cause of stereotype generation through direct neural intervention. Our solution introduces two synergistic components: Diffusion Deep Taylor Decomposition (DDTD) for precisely localizing stereotype-related neurons via Layer-wise Relevance Propagation (LRP) attribution analysis, and Stereotype Neuron Suppression (SNS) implementing targeted activation damping to neutralize bias propagation. Through extensive empirical evaluations across multiple bias dimensions, we demonstrate that our method achieves significant stereotype mitigation without compromising image quality or requiring additional training data. This neuro-inspired approach establishes a new paradigm for model interpretability and ethical alignment in generative AI systems.

Yan Yang, Zeguan Xiao, Xin Lu, Hongru Wang, Xuetao Wei, Hailiang Huang, Guanhua Chen, Yun Chen. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Building a universal trajectory foundation model is a promising solution to address the limitations of existing trajectory modeling approaches, such as task specificity, regional dependency, and data sensitivity. Despite its potential, data preparation, pre-training strategy development, and architectural design present significant challenges in constructing this model. Therefore, we introduce **UniTraj**, a Universal Trajectory foundation model that aims to address these limitations through three key innovations. First, we construct **WorldTrace**, an unprecedented dataset of 2.45 million trajectories with billions of GPS points spanning 70 countries, providing the diverse geographic coverage essential for region-independent modeling. Second, we develop novel pre-training strategies--Adaptive Trajectory Resampling and Self-supervised Trajectory Masking--that enable robust learning from heterogeneous trajectory data with varying sampling rates and quality. Finally, we tailor a flexible model architecture to accommodate a variety of trajectory tasks, effectively capturing complex movement patterns to support broad applicability. Extensive experiments across multiple tasks and real-world datasets demonstrate that UniTraj consistently outperforms existing methods, exhibiting superior scalability, adaptability, and generalization, with WorldTrace serving as an ideal yet non-exclusive training resource. The implementation codes and full dataset are available at https://github.com/Yasoz/UniTraj.

Recent deep sequential recommendation models often struggle to effectively model key characteristics of user behaviors, particularly in handling sequence length variations and capturing diverse interaction patterns. We propose STAR-Rec, a novel architecture that synergistically combines preference-aware attention and state-space modeling through a sequence-level mixture-of-experts framework. STAR-Rec addresses these challenges by: (1) employing preference-aware attention to capture both inherently similar item relationships and diverse preferences (2) utilizing state-space modeling to efficiently process variable-length sequences with linear complexity, and (3) incorporating a mixture-of-experts component that adaptively routes different behavioral patterns to specialized experts, handling both focused category-specific browsing and diverse category exploration patterns. We theoretically demonstrate how the state space model and attention mechanisms can be naturally unified in recommendation scenarios, where SSM captures temporal dynamics through state compression while attention models both similar and diverse item relationships. Extensive experiments on four real-world datasets demonstrate that STAR-Rec consistently outperforms state-of-the-art sequential recommendation methods, particularly in scenarios involving diverse user behaviors and varying sequence lengths. The implementation code is available anonymously online for easy reproducibility.

Delivering superior search services is crucial for enhancing customer experience and driving revenue growth in e-commerce. Conventionally, search systems model user behaviors by combining user preference and query-item relevance statically, often through a fixed logical 'and' relationship. This paper reexamines existing approaches through a unified lens using causal graphs and Venn diagrams, uncovering two prevalent yet significant issues: entangled preference and relevance effects, and a collapsed modeling space. To surmount these challenges, our research introduces a novel framework, DRP, which enhances search accuracy through two components to reconstruct the behavior modeling space. Specifically, we implement preference editing to proactively remove the relevance effect from preference predictions, yielding untainted user preferences. Additionally, we employ adaptive fusion, which dynamically adjusts fusion criteria to align with the varying patterns of relevance and preference, facilitating more nuanced and tailored behavior predictions within the reconstructed modeling space. Empirical validation on two public datasets and a proprietary e-commerce search dataset underscores the superiority of our proposed methodology, demonstrating marked improvements in performance over existing approaches. The code is available at https://github.com/Applied-Machine-Learning-Lab/DRP.

Affective robots can elicit psychological responses such as attachment and intimacy, which may help alleviate anxiety and enrich users’ emotional experiences. While such robots can take the form of agents (controlled by algorithms) or avatars (controlled by humans or animals), the differential effects of these embodiment forms on users’ emotional responses remain underexplored, particularly in scenarios where avatars are controlled by animals. In this study, we conducted a Wizard of Oz experiment to compare the emotional experience and anxiety relief provided by a robotic cat under different embodiment forms and interaction modes (unidirectional vs. bidirectional). The results indicate that the avatar embodiment significantly enhances users’ affective experiences, fostering stronger emotional bonds and more effective anxiety relief. However, no significant differences were found between the interaction modes with respect to either anxiety relief or emotional outcomes. These findings offer valuable insights for the design of emotionally engaging embodied intelligent systems in contexts such as emotional companionship and mental health interventions. • A wearable robot prototype enables remote interaction between humans and real cats. • Robot embodiment forms and interaction modes affect users’ anxiety and affective experience. • The avatar form leads to greater anxiety relief and stronger affective experiences. • No significant emotional differences are found between interaction modes.

In sharded blockchains, peers are divided into smaller groups (shards) that generate and verify blocks in parallel, offering enhanced throughput and reduced delays. These properties make sharded blockchains a promising solution for secure data management in Internet of Things (IoT) systems. Particularly, geolocation-based sharded blockchains assign geographically proximate peers to the same shard, enabling faster IoT transaction processing. Yet, peers in each shard can easily collude to falsely accept/reject blocks. To resolve this issue, in this paper, we propose a robust reputation-based shard inspection framework. The framework adopts the shard inspection mechanism where a group of inspectors selected from the most reputable peers randomly verify blocks in each shard. This enables avoiding collusion attacks and enhancing the security of each shard. However, additional block verifications during the inspection process can incur significant block delays and energy overheads. To reduce these overheads, we formulate an optimization problem that jointly determines the number of inspectors and the inspection interval to maximize the system utility, which is proportional to the blockchain throughput and energy consumption. We then develop a distributed algorithm that enables dividing the optimization problem into sub-problems solvable independently by each shard. Experimental results show that our framework can maximize the system utility, while maintaining high levels of security in each shard.

We study decentralized federated learning (DFL) in edge computing networks where edge nodes (ENs) collaboratively train their artificial intelligence (AI) models in a serverless manner without sharing local data. We consider the following critical DFL challenges: i) scarce bandwidth resources of ENs; ii) dynamic, heterogeneous edge environment; iii) incentive provisioning and complex tradeoffs between the DFL performance and training costs. To resolve these challenges, we develop a new model compression method where ENs utilize dynamic, non-identical compression rates to improve the communication efficiency of DFL under time-varying, heterogeneous resource constraints. We show that our method can be formulated as a graphical Markov potential game where ENs act as players deciding on their compression factors and the number of data samples used for model updates. Each EN is incentivized to participate in DFL through rewards based on the EN's contribution to training. We prove that our game has a dominant pure-strategy Nash equilibrium (NE) maximizing its potential function and propose a dynamic distributed compression algorithm in which each EN can find its dominant strategy independently. We show that this algorithm converges to the Pareto-optimal NE, representing the most efficient solution of our game enhancing the DFL performance with minimal costs.

The goal of 3D mesh watermarking is to embed the message in 3D meshes that can withstand various attacks imperceptibly and reconstruct the message accurately from watermarked meshes. The watermarking algorithm is supposed to withstand multiple attacks, and the complexity should not grow significantly with the mesh size. Unfortunately, previous methods are less robust against attacks and lack of adaptability. In this paper, we propose a robust and adaptable deep 3D mesh watermarking Deep3DMark that leverages attention-based convolutions in watermarking tasks to embed binary messages in vertex distributions without texture assistance. Furthermore, our Deep3DMark exploits the property that simplified meshes inherit similar relations from the original ones, where the relation is the offset vector directed from one vertex to its neighbor. By doing so, our method can be trained on simplified meshes but remains effective on large size meshes (size adaptable) and unseen categories of meshes (geometry adaptable). Extensive experiments demonstrate our method remains efficient and effective even if the mesh size is 190× increased. Under mesh attacks, Deep3DMark achieves 10%∼50% higher accuracy than traditional methods, and 2× higher SNR and 8% higher accuracy than previous DNN-based methods.

The online data markets have emerged as a valuable source of diverse datasets for training machine learning (ML) models. However, datasets from different data providers may exhibit varying levels of bias with respect to certain sensitive attributes in the population (such as race, sex, age, and marital status). Recent dataset acquisition research has focused on maximizing accuracy improvements for downstream model training, ignoring the negative impact of biases in the acquired datasets, which can lead to an unfair model. Can a consumer obtain an unbiased dataset from datasets with diverse biases? In this work, we propose a fairness-aware data acquisition framework (FAIRDA) to acquire high-quality datasets that maximize both accuracy and fairness for consumer local classifier training while remaining within a limited budget. Given the biases of data commodities remain opaque to consumers, the data acquisition in FAIRDA employs explore-exploit strategies. Based on whether exploration and exploitation are conducted sequentially or alternately, we introduce two algorithms: the knowledge-based offline data acquisition (KDA) and the reward-based online data acquisition algorithms (RDA). Each algorithm is tailored to specific customer needs, giving the former an advantage in computational efficiency and the latter an advantage in robustness. We conduct experiments to demonstrate the effectiveness of the proposed data acquisition framework in steering users toward fairer model training compared to existing baselines under varying market settings.

Recommender systems play a pivotal role across practical scenarios, showcasing remarkable capabilities in user preference modeling. However, the centralized learning paradigm predominantly used raises serious privacy concerns. The federated recommender system (FedRS) addresses this by updating models on clients, while a central server orchestrates training without accessing private data. Existing FedRS approaches, however, face unresolved challenges, including non-convex optimization, vulnerability, potential privacy leakage risk, and communication inefficiency. This paper addresses these challenges by reformulating the federated recommendation problem as a convex optimization issue, ensuring convergence to the global optimum. Based on this, we devise a novel method, RFRec, to tackle this optimization problem efficiently. In addition, we propose RFRecF, a highly efficient version that incorporates non-uniform stochastic gradient descent to improve communication efficiency. In user preference modeling, both methods learn local and global models, collaboratively learning users' common and personalized interests under the federated learning setting. Moreover, both methods significantly enhance communication efficiency, robustness, and privacy protection, with theoretical support. Comprehensive evaluations on four benchmark datasets demonstrate RFRec and RFRecF's superior performance compared to diverse baselines. The code is available to ease reproducibility1.

Federated learning (FL) stands as a paradigmatic approach that facilitates model training across heterogeneous and diverse datasets originating from various data providers. However, conventional FLs fall short of achieving consistent performance, potentially leading to performance degradation for clients who are disadvantaged in data resources. Influenced by the Matthew effect, deploying a performance-imbalanced global model in applications further impedes the generation of high-quality data from disadvantaged clients, exacerbating the disparities in data resources among clients. In this work, we propose anti-Matthew fairness for the global model at the client level, requiring equal accuracy and equal decision bias across clients. To balance the trade-off between achieving anti-Matthew fairness and performance optimality, we formalize the anti-Matthew effect federated learning (anti-Matthew FL) as a multi-constrained multi-objectives optimization (MCMOO) problem and propose a three-stage multi-gradient descent algorithm to obtain the Pareto optimality. We theoretically analyze the convergence and time complexity of our proposed algorithms. Additionally, through extensive experimentation, we demonstrate that our proposed anti-Matthew FL outperforms other state-of-the-art FL algorithms in achieving a high-performance global model while effectively bridging performance gaps among clients. We hope this work provides valuable insights into the manifestation of the Matthew effect in FL and other decentralized learning scenarios and can contribute to designing fairer learning mechanisms, ultimately fostering societal welfare.

Deep Recommender Systems (DRS) are essential for navigating the extensive data across various platforms in today's digital landscape. Current DRS models often treat all features equally and implement complex structures to enhance the capture of feature interactions. However, they may fail to recognize crucial user patterns due to not fully utilizing user-specific features for user modeling. Moreover, prevailing user modeling techniques concentrate exclusively on either the group or individual level, overlooking the potential insights from the unaddressed one. This oversight can miss shared group preferences or learn group patterns that conflict with individual preferences. To overcome these limitations, we introduce GPRec, a novel bi-level user modeling approach that substantially improves DRS. GPRec explicitly categorizes users into groups in a learnable manner and aligns them with corresponding group embeddings. We design the dual group embedding space to offer a diverse perspective on group preferences by contrasting positive and negative patterns. On the individual level, GPRec identifies personal preferences from ID-like features and refines the obtained individual representations to be independent of group ones, thereby providing a robust complement to the group-level modeling. We also present various strategies for the flexible integration of GPRec into various DRS models. Rigorous testing of GPRec on three public datasets has demonstrated significant improvements in recommendation quality. Additional experiments further explore crucial components of GPRec, its parameter sensitivity, and the group diversity. The implementation code is readily available online to facilitate future research and practical deployment: https://github.com/Applied-Machine-Learning-Lab/GPRec.

Generating trajectory data is among promising solutions to addressing privacy concerns, collection costs, and proprietary restrictions usually associated with human mobility analyses. However, existing trajectory generation methods are still in their infancy due to the inherent diversity and unpredictability of human activities, grappling with issues such as fidelity, flexibility, and generalizability. To overcome these obstacles, we propose ControlTraj, a Controllable Trajectory generation framework with the topology-constrained diffusion model. Distinct from prior approaches, ControlTraj utilizes a diffusion model to generate high-fidelity trajectories while integrating the structural constraints of road network topology to guide the geographical outcomes. Specifically, we develop a novel road segment autoencoder to extract fine-grained road segment embedding. The encoded features, along with trip attributes, are subsequently merged into the proposed geographic denoising UNet architecture, named GeoUNet, to synthesize geographic trajectories from white noise. Through experimentation across three real-world data settings, ControlTraj demonstrates its ability to produce human-directed, high-fidelity trajectory generation with adaptability to unexplored geographical contexts.

DNN-based watermarking methods are rapidly developing and delivering impressive performances. Recent advances achieve resolution-agnostic image watermarking by reducing the variant resolution watermarking problem to a fixed resolution watermarking problem. However, such a reduction process can potentially introduce artifacts and low robustness. To address this issue, we propose the first, to the best of our knowledge, Resolution-Agnostic Image WaterMarking (RAIMark) framework by watermarking the implicit neural representation (INR) of image. Unlike previous methods, our method does not rely on the previous reduction process by directly watermarking the continuous signal instead of image pixels, thus achieving resolution-agnostic watermarking. Precisely, given an arbitrary-resolution image, we fit an INR for the target image. As a continuous signal, such an INR can be sampled to obtain images with variant resolutions. Then, we quickly fine-tune the fitted INR to get a watermarked INR conditioned on a binary secret message. A pre-trained watermark decoder extracts the hidden message from any sampled images with arbitrary resolutions. By directly watermarking INR, we achieve resolution-agnostic watermarking with increased robustness. Extensive experiments show that our method outperforms previous methods with significant improvements: averagely improved bit accuracy by 7%~29%. Notably, we observe that previous methods are vulnerable to at least one watermarking attack (e.g. JPEG, Crop, or Resize), while ours are robust against all watermarking attacks.

Federated learning (FL) offers a machine learning paradigm that protects privacy, allowing multiple clients to collaboratively train a global model while only accessing their local data. Recent research in FL has increasingly focused on improving the uniformity of model performance across clients, a fairness principle known as egalitarian fairness. However, achieving egalitarian fairness in FL may sacrifice the model performance for data-rich clients to benefit those with less data. This trade-off raises concerns about the stability of FL, as data-rich clients may opt to leave the current coalition and join another that is more closely aligned with its expected high performance. In this context, our work rigorously addresses the critical concern: **Does egalitarian fairness lead to instability?** Drawing from game theory and social choice theory, we initially characterize fair FL systems as altruism coalition formation games (ACFGs) and reveal that the instability issues emerging from the pursuit of egalitarian fairness are significantly related to the clients’ altruism within the coalition and the configuration of the friends-relationship networks among the clients. Then, we theoretically propose the optimal egalitarian fairness bounds that an FL coalition can achieve while maintaining core stability under various types of altruistic behaviors. The theoretical contributions clarify the quantitative relationships between achievable egalitarian fairness and the disparities in the sizes of local datasets, disproving the misconception that egalitarian fairness inevitably leads to instability. Finally, we conduct experiments to evaluate the consistency of our theoretically derived egalitarian fairness bounds with the empirically achieved egalitarian fairness in fair FL settings.

Worldwide geolocalization aims to locate the precise location at the coordinate level of photos taken anywhere on the Earth. It is very challenging due to 1) the difficulty of capturing subtle location-aware visual semantics, and 2) the heterogeneous geographical distribution of image data. As a result, existing studies have clear limitations when scaled to a worldwide context. They may easily confuse distant images with similar visual contents, or cannot adapt to various locations worldwide with different amounts of relevant data. To resolve these limitations, we propose **G3**, a novel framework based on Retrieval-Augmented Generation (RAG). In particular, G3 consists of three steps, i.e., **G**eo-alignment, **G**eo-diversification, and **G**eo-verification to optimize both retrieval and generation phases of worldwide geolocalization. During Geo-alignment, our solution jointly learns expressive multi-modal representations for images, GPS and textual descriptions, which allows us to capture location-aware semantics for retrieving nearby images for a given query. During Geo-diversification, we leverage a prompt ensembling method that is robust to inconsistent retrieval performance for different image queries. Finally, we combine both retrieved and generated GPS candidates in Geo-verification for location prediction. Experiments on two well-established datasets IM2GPS3k and YFCC4k verify the superiority of G3 compared to other state-of-the-art methods. Our code is available online [https://github.com/Applied-Machine-Learning-Lab/G3](https://github.com/Applied-Machine-Learning-Lab/G3) for reproduction.

Text-to-Image (T2I) has witnessed significant advancements, demonstrating superior performance for various generative tasks. However, the presence of stereotypes in T2I introduces harmful biases that require urgent attention as the T2I technology becomes more prominent. Previous work for stereotype mitigation mainly concentrated on mitigating stereotypes engendered with individual objects within images, which failed to address stereotypes engendered by the association of multiple objects, referred to as *Association-Engendered Stereotypes*. For example, mentioning ''black people'' and ''houses'' separately in prompts may not exhibit stereotypes. Nevertheless, when these two objects are associated in prompts, the association of ''black people'' with ''poorer houses'' becomes more pronounced. To tackle this issue, we propose a novel framework, MAS, to Mitigate Association-engendered Stereotypes. This framework models the stereotype problem as a probability distribution alignment problem, aiming to align the stereotype probability distribution of the generated image with the stereotype-free distribution. The MAS framework primarily consists of the *Prompt-Image-Stereotype CLIP* (*PIS CLIP*) and *Sensitive Transformer*. The *PIS CLIP* learns the association between prompts, images, and stereotypes, which can establish the mapping of prompts to stereotypes. The *Sensitive Transformer* produces the sensitive constraints, which guide the stereotyped image distribution to align with the stereotype-free probability distribution. Moreover, recognizing that existing metrics are insufficient for accurately evaluating association-engendered stereotypes, we propose a novel metric, *Stereotype-Distribution-Total-Variation*(*SDTV*), to evaluate stereotypes in T2I. Comprehensive experiments demonstrate that our framework effectively mitigates association-engendered stereotypes.

Large Language Models (LLMs) have demonstrated remarkable capabilities, surpassing human experts in various benchmark tests and playing a vital role in various industry sectors. Despite their effectiveness, a notable drawback of LLMs is their inconsistent moral behavior, which raises ethical concerns. This work delves into symmetric moral consistency in large language models and demonstrates that modern LLMs lack sufficient consistency ability in moral scenarios. Our extensive investigation of twelve popular LLMs reveals that their assessed consistency scores are influenced by position bias and selection bias rather than their intrinsic abilities. We propose a new framework tSMC, which gauges the effects of these biases and effectively mitigates the bias impact based on the Kullback–Leibler divergence to pinpoint LLMs' mitigated Symmetric Moral Consistency. We find that the ability of LLMs to maintain consistency varies across different moral scenarios. Specifically, LLMs show more consistency in scenarios with clear moral answers compared to those where no choice is morally perfect. The average consistency score of 12 LLMs ranges from $60.7\%$ in high-ambiguity moral scenarios to $84.8\%$ in low-ambiguity moral scenarios.

Exploiting spatial–temporal correlations has long been regarded as the cornerstone of traffic state prediction. Among existing techniques, temporal graph neural networks (TGNNs) have recently emerged as a prominent solution for modeling complex spatial–temporal traffic data correlations. Existing studies on TGNNs mainly focus on developing new building blocks to embed hidden correlations into a unified latent representation, which is mapped to predictions of distinct horizons. However, mapping the same latent features to distinct scalar predictions makes the gradient computation challenging for updating model parameters in the relevant directions. Besides, TGNNs are biased towards the shared temporal patterns while neglecting the complex dependencies within each data series, which can be captured to enrich latent features. To handle these problems jointly, we propose a novel feature projection scheme for the traffic prediction framework of TGNNs. The proposed projection scheme is based on spatial convolutions that first generate horizon-specific feature maps and then transform them into scalar predictions of the corresponding horizons. These horizon-specific feature maps establish interactions between the unified latent representation and the corresponding output values to bring the predictions closer to the true values. Besides, the proposed scheme also serves as a pattern modeling phase that enhances the expressivity of TGNNs by enriching latent features with data source-wise patterns of distinct time steps. Comprehensive experiments on two real-world traffic datasets demonstrate that the proposed scheme enhances the predictive performance and reduces the model parameters of TGNNs.

We study the problem of ultra-reliable and low-latency slicing in multi-access edge computing (MEC) systems for the next-generation internet of things (IoT) and mobile applications operating in the space-air-ground integrated network. The network has a dynamic topology formed by multiple non-stationary nodes with unstable communication links and unreliable processing/transmission resources. Each node can be in one of two hidden states: i) reliable – in which the node generates no data errors and no losses; ii) unreliable – when the node can generate/propagate random data errors/losses. Solving this problem is difficult, as it represents the non-deterministic polynomial-time (NP) hard non-concave non-smooth stochastic maximization problem which depends on the unknown hidden nodes’ states and private information about local, dynamic parameters of each node, which is known only to this node, and not to other nodes. To address these challenges, we develop a new deep learning (DL) model based on the message passing graph neural network (MPNN) to estimate hidden nodes’ states. We then propose a novel algorithm based on the online alternating direction method of multipliers (ADMM) – an extension of the well-known classical “static” ADMM to dynamic settings, where our slicing problem can be solved distributedly, in real time, without revealing local (private) information of the nodes. We show that our algorithm converges to a global optimum of the slicing problem and has a good consistent performance even in highly-dynamic, unreliable scenarios.

Programmable smart contracts facilitate the application of blockchain to many areas of industry, e.g., supply chains, energy, and healthcare. The off-chain execution model for blockchain smart contracts is a promising approach to achieve scalability for industrial applications, which executes each smart contract in one of the small-scale consensus groups. However, previous work based on such a model cannot efficiently handle cross-group contract transactions that involve multiple cross-group invocations between smart contracts because of the heavyweight state synchronization and repeated executions. In this article, we propose interactive consensus-based off-chain execution (ICOE), a lightweight group-consensus-based off-chain execution model, for cross-group contract transactions without the need of state synchronization and duplicate executions and also prove its correctness. The key novelty of ourICOEis to let the invoked smart contract only executed in its residing group and only return the execution result to the invoking smart contract while leveraging the two-phase commit protocol to guarantee consistent commits across multiple groups within the same invoking smart contract transaction. We implementICOEand the extensive experimental results show that ourICOEachieves a 14× improvement of throughput on average compared with the state-of-the-art model.

We study the problem of distributed computing in themulti-operator multi-access edge computing(MEC) network fordependent tasks. Every task comprises severalsub-taskswhich are executed based on logical precedence modelled as adirected acyclic graph. In the graph, each vertex is a sub-task, each edge – precedence constraint, such that a sub-task can only be started after all its preceding sub-tasks are completed. Tasks are executed by MEC servers with the assistance of nearby edge devices, so that the MEC network can be viewed as adistributed“primary-secondary node” system where each MEC server acts as aprimary node(PN) deciding on sub-tasks assigned to itssecondary nodes(SNs), i.e., nearby edge devices. The PN's decision problem is complex, as its SNs can be associated with otherneighboringPNs. In this case, the available processing resources of SNs depend on the sub-task assignment decisions of all neighboring PNs. Since PNs are controlled by different operators, they do not coordinate their decisions, and each PN is uncertain about the sub-task assignments of its neighbors (and, thus, the available resources of its SNs). To address this problem, we propose a novel framework based on agraphical Bayesian game, where PNs play under uncertainty about their neighbors’ decisions. We prove that the game has aperfect Bayesian equilibrium(PBE) yieldingunique optimal values, and formulate newBayesian reinforcement learningandBayesian deep reinforcement learningalgorithms enabling each PN to reach the PBE autonomously (without communicating with other PNs).

Concurrent execution of smart contracts is a promising approach to boost their performance in current blockchain systems. However, prior work on concurrent execution only considered each smart contract transaction as a single atomic unit for the concurrent scheduling, which is not suitable for the increasingly complex smart contracts with high contention in practice. In this article, we presentSC-Chef, a new concurrency control system to turboboost smart contract concurrent execution for high contention workloads via the transaction decomposition. The key novelty of ourSC-Chefis to decompose the original transaction into multiple subtransactions that can be executed independently, while still respecting the original logic dependency. We implement and evaluate ourSC-Chefin the private Ethereum platform. Our extensive experiments show that ourSC-Chefoutperforms the classic optimistic concurrency control and two-phase lock by 60% and 70% on average in the case of high contention, respectively.

Consortium blockchains have attracted considerable interest from academia and industry due to their low-cost installation and maintenance. However, typical consortium blockchains can be easily attacked by colluding block validators because of the limited number of miners in the systems. To address this problem, in this paper, we propose a novel block validation framework to enhance blockchain security. In the framework, the block validations are assisted and implemented by various lightweight nodes, e.g., edge devices, in addition to the typical blockchain miners. This improves the blockchain security but can cause an increased block validation delay and, thereby, reduced blockchain throughput. To tackle this challenge, we propose an effective method to select lightweight nodes based on their computing powers to maximize the blockchain throughput, and prove the uniqueness of the optimal nodes selection strategy. Security analysis and simulation results from the deployed consortium blockchain platform show that the proposed framework achieves higher throughput and security than the existing consortium blockchain models.

Most NFTs (Non-Fungible Tokens) use multi-hop URLs to address the off-chain assets due to the costly on-chain storage, but the path from NFTs to the underlying assets is fraught with instability, which may degrade its value. Hence, this paper aims to answer the question: Is the NFT-to-Asset connection fragile? This paper makes a first step towards this end by characterizing NFT-to-Asset connections of 12,353 Ethereum NFT Contracts (6,234,141 NFTs in total) from three perspectives, storage, accessibility, and duplication. In order to overcome challenges of affecting the measurement accuracy, e.g., IPFS instability and the changing availability of both IPFS and servers’ data, we propose to leverage multiple gateways to enlarge the data coverage and extend a longer measurement period with non-trivial efforts. Results of our extensive study show that such connection is very fragile in practice. The loss, unavailability, or duplication of off-chain assets could render the value of NFTs worthless. For instance, we find that assets of 25.24% of Ethereum NFT contracts are not accessible, and 21.48% of Ethereum NFT contracts include duplicated assets. Our work sheds light on the fragility along the NFT-to-Asset connection, which could help the NFT community to better enhance the trust of off-chain assets.

State sharding is a promising approach to address the scalability issue in the blockchain system. However, the previous sharding schemes using the fixed data partitioning mechanism bring high proportion of costly cross-shard transactions and cannot effectively handle the workload imbalance that occurs in practice, which slows down the performance. To address these issues, we propose EFSHARD, an efficient state sharding blockchain system that enables allocating states flexibly and timely across shards according to recent transactions. Firstly, we propose a hierarchical state partition to enable flexible mapping of states to shards. Second, we design a new state transfer protocol to efficiently migrate states across shards while guaranteeing consistency and liveness. Then, we provide a greedy-based state allocation algorithm to decide when and how to migrate states. The allocation mechanism groups highly correlated state data into the same shard to reduce the proportion of cross-shard transactions and distributes state data to shards with relatively low load to balance workload, thus improving the performance. In the end, we conduct extensive experiments to evaluate EFSHARD and the results demonstrate that EFSHARD outperforms state-of-the-art approaches in terms of transaction throughput, confirmation latency, workload balance, and queue size of transaction pool.

The data query service is urgently required in sustainable-storage blockchain, where full nodes store the entire transaction data while light nodes only store block headers. Queries invariably seek data with multiple attributes. However, no existing method provides a unified authenticated data structure (ADS) to support complex query operators (e.g., range queries and data object queries) on multiattribute blockchain data. In this paper, we propose a framework EAQ that effectively supports both fast data queries on multiple attributes and authentication of the query result. We propose a new ADS, called the MR$^{Bloom}$-tree, based on the Bloom filter (BF) and Merkle R-tree. We prove the decomposability of BFs, which enables the BF to be seamlessly incorporated with the Merkle R-tree. This ADS enables range-level search using multidimensional attribute ranges and object-level search using BFs. This ADS also supports querying and proving inexistent data objects. To reduce storage overhead, we improve the MR$^{Bloom}$-tree using the suppressed BF structure, which constructs only one BF independent of the number of attributes. To manage string attributes, we transform them into discrete numerical attributes using density-based clustering to represent similar items with close numerical values. Experiments show that the proposed framework achieves promising results.

Visible-light positioning (VLP) has been considered as a promising indoor positioning technology due to its high precision and low cost. However, current VLP techniques are greatly limited by the assumption that lights have to be turned on to emit shining light beams, which is not applicable to scenarios that do not need the illumination all the time. In this article, we design and implement a novel VLP system, DarkVLP, to achieve high-precision positioning when lights are “turned off.” The “turn off” state means lights emit the extremely low luminance that is imperceptible to human eyes. In order to realize such a system, we have to tackle nontrivial challenges in data encoding/decoding, modulation/demodulation, and positioning algorithm. Specifically, we propose the sliding rheostat-based modulation scheme to eliminate the need of a complex signal synchronization mechanism and landmark recognition algorithm at the receiver. We also propose the dual-photodiode-based positioning algorithm, which effectively mitigates effects of substantial signal strength fluctuation, to reliably achieve VLP. In the end, we design novel circuits and prototype our system DarkVLP on commercial off-the-shelf devices. The results of extensive experiments demonstrate that our DarkVLP system could achieve submeter precision under the extremely low luminance, which greatly broadens application scenarios of VLP.

Big data and cognitive computing have a wide range of applications in smart homes, smart cities, artificial intelligence, and computational social systems. Visible light positioning systems have attracted more and more attention as one of the application scenarios of computational social systems. In visible light positioning systems, the light-emitting diodes (LEDs) ceiling layout makes the smartphone usually obtain less than three LEDs in captured images. Due to the lack of necessary positioning information, most scholars combine the inertial measurement unit (IMU) with the modified filter algorithms to achieve positioning under the sparse light source. However, these systems have the following problems: 1) the azimuth angle obtained by the IMU is always not accurate, which decreases the positioning accuracy and 2) during the dynamic positioning process, the system’s initial position is difficult to automatically determine. In this article, we propose indoor high-precision visible light positioning under the sparse light source. First, we propose the geometric correction mechanism, which uses ellipse fitting to calibrate the azimuth angle, so as to increase the positioning accuracy for the static system. Then, we build a motion model for the entire positioning process through the unscented particle filter (UPF), which does not need to manually set initial state parameters, due to random generated particles. It can increase the positioning accuracy for the dynamic system. We evaluate our designed system, and the experimental results show that the average azimuth angle error is 2.04° and average positioning error is 8.8 cm, under the sparse light source.

The proliferation of the Internet of Things pushes the visible light positioning (VLP) system research. However, the existing positioning systems still have the following problems: 1) when the smartphone (receiver) is rotated or tilted during the positioning, existing collaborative LEDs’ positioning algorithms fail and 2) for different smartphone application scenarios, there is not an effective resource management solution between the server and the client. Therefore, in this article, we design and implement a robust and flexible indoor VLP system based on collaborative LEDs and edge computing. First, we propose the enhanced collaborative LEDs’ positioning algorithm, which uses the indoor hidden location information to obtain the rotation angle and tilt angle of the receiver, to achieve the robust system positioning. Then, we use the edge computing solution to balance between bandwidth resources and computing resources and propose a flexible functional segmentation scheme for different smartphone application scenarios. Finally, we conduct experimental tests to evaluate the positioning system performance by landmark decoding rate, positioning accuracy, and segmentation analysis. Test results show that the designed positioning system can achieve centimeter-level positioning. Meanwhile, the smartphone can exchange the least bandwidth resources for the most computing resources under Scheme-3.

With the great development in smart devices, mobile crowdsensing (MCS) has been an innovative paradigm for data gathering. Task assignment is a fundamental problem in MCS systems and applications. Previous studies only focused on the assignment of individual tasks, neglecting planning the task processing from a higher level, e.g., making assignments between task locations and workers, which impacts the crowdsensing performance adversely. Furthermore, task characteristics, e.g., route distance, task similarity, and task priority, have a great impact on the rationality of a visiting order and the quality of crowdsensing services. In this article, we tackle the problem of rational task assignment and path planning for MCS, which aims to assign a set of task locations to a set of workers and generate location visiting sequences. We measure the assignment rationality by taking into account geographical information and task characteristics, i.e., route distance, task similarity, and task priority. We prove that the problem of computing the rationality maximization task assignment is NP-hard. For the single-worker scenario, we reduce the problem to a simpler problem with respect to only route distance and task priority criteria since the similarity measurement has a fixed value. We propose an effective greedy algorithm. For the multiple-worker scenario, we first extend the greedy idea by considering all three criteria and then propose an effective approach by reducing the computational complexity of similarity measurement. Extensive experiments show that our proposed approaches achieve promising results.

The heavy storage problem has become a key obstacle to the application of blockchain to the actual business environments, because each node needs to keep a complete replica of blockchain data. The data volume grows undesirably large in practice. It prevents the widely used devices, e.g., tablets and mobile phones, to join blockchain systems due to their limited storage and computing resource. Previous work addressed the storage issue by allowing participating nodes to only keep a fraction of the entire transaction set, e.g., sharding. However, existing studies focus on transaction placement with the minimum cross-shard communications. These studies neglect the node characteristics (e.g., storage capacity, cost, and response capability), which impacts the storage performance adversely. In this paper, we propose EBSF, a block storage framework that achieves efficient block storage by constructing a block allocation plan based on node characteristics. Blockchain nodes are organized into committees such that nodes in a committee work together to maintain the entire blockchain data. We formulate the block allocation plan problem that assigns each block to at least one node in a committee. The goal is to minimize the total cost while reaching the threshold of the response capability of each block. We prove the NP-hardness of the problem and propose heuristic algorithms. We also propose two strategies to handle the dynamic scenario of new blocks. Extensive evaluation shows the efficiency and effectiveness of the proposed framework.

The time series data generated by massive sensors in Internet of Things (IoT) is extremely dynamic, heterogeneous, large scale and time-dependent. It poses great challenges (e.g. accuracy, reliability, stability) on the real-time analysis and decision making for different IoT applications. In this paper, we design, implement and evaluate EdgeLSTM, a unified data-driven system to enhance IoT computing at the network edge. The EdgeLSTM leverages the grid long short-term memory (Grid LSTM) to provide an agile solution for both deep and sequential computation, therefore can address important features such as large-scale, variety, time dependency and real time in IoT data. Our system exploits the advantages of Grid LSTM network and extends it with a multiclass support vector machine by rigorous regularization and optimization approaches, which not only has strong prediction capability of time series data, but also achieves fine-grained multiple classification through the predictive error. We deploy the EdgeLSTM into four IoT applications, including data prediction, anomaly detection, network maintenance and mobility management by extensive experiments. Our evaluation results of real-world time series data with different short-term and long-term time dependency from these typical IoT applications show that our EdgeLSTM system can guarantee robust performance in IoT computing.

With the rapid development of smart cities, white light‐emitting diodes (LEDs) are widely used in indoor lighting due to their characteristics of energy‐saving, long lifetimes, and low cost. At the same time, the high‐frequency modulated LEDs allow visible light to be used for indoor communications, positioning, and security. Visible light applications have many advantages, such as avoiding electromagnetic interference, high communication speed, great privacy, and high‐precise positioning. In this paper, we will survey the application prospects and research results of visible light from scenarios such as high‐speed communication, privacy security, and navigation localization and focus on analyzing and discussing the challenges and development trends encountered in visible light positioning.

Dynamic Information Flow Tracking (DIFT), also called Dynamic Taint Analysis (DTA), is a technique for tracking the information as it flows through a program’s execution. Specifically, some inputs or data get tainted and then these taint marks (tags) propagate usually at the instruction-level. While DIFT has been a fundamental concept in computer and network security for the past decade, it still faces open challenges that impede its widespread application in practice; one of them being the indirect flow propagation dilemma: should the tags involved in an indirect flow, e.g., in a control or address dependency, be propagated? Propagating all these tags, as is done for direct flows, leads to overtainting (all taintable objects become tainted), while not propagating them leads to undertainting (information flow becomes incomplete). In this paper, we analytically model that decisioning problem for indirect flows, by considering various tradeoffs including undertainting versus overtainting, importance of heterogeneous code semantics and context. Towards tackling this problem, we design MITOS, a distributed-optimization algorithm, that: decides about the propagation of indirect flows by properly weighting all these tradeoffs, is of low-complexity, is scalable, is able to flexibly adapt to different application scenarios and security needs of large distributed systems. Additionally, MITOS is applicable to most DIFT systems that consider an arbitrary number of tag types, and introduces the key properties of fairness and tag-balancing to the DIFT field. To demonstrate MITOS’s applicability in practice, we implement and evaluate MITOS on top of an open-source DIFT, and we shed light on the open problem. We also perform a case-study scenario with a real in-memory only attack and show that MITOS improves simultaneously (i) system’s spatiotemporal overhead (up to 40%), and (ii) system’s fingerprint on suspected bytes (up to 167%) compared to traditional DIFT, even though these metrics usually conflict.

This paper investigates the privacy leakage of smart speakers under an encrypted traffic analysis attack, referred to as voice command fingerprinting. In this attack, an adversary can eavesdrop both outgoing and incoming encrypted voice traffic of a smart speaker, and infers which voice command a user says over encrypted traffic. We first built an automatic voice traffic collection tool and collected two large-scale datasets on two smart speakers, Amazon Echo and Google Home. Then, we implemented proof-of-concept attacks by leveraging deep learning. Our experimental results over the two datasets indicate disturbing privacy concerns. Specifically, compared to 1% accuracy with random guess, our attacks can correctly infer voice commands over encrypted traffic with 92.89% accuracy on Amazon Echo.

The data generated by millions of sensors in the industrial Internet of Things (IIoT) are extremely dynamic, heterogeneous, and large scale and pose great challenges on the real-time analysis and decision making for anomaly detection in the IIoT. In this article, we propose a long short-term memory (LSTM)-Gauss-NBayes method, which is a synergy of the long short-term memory neural network (LSTM-NN) and the Gaussian Bayes model for outlier detection in the IIoT. In a nutshell, the LSTM-NN builds a model on normal time series. It detects outliers by utilizing the predictive error for the Gaussian Naive Bayes model. Our method exploits advantages of both LSTM and Gaussian Naive Bayes models, which not only has strong prediction capability of LSTM for future time point data, but also achieves an excellent classification performance of the Gaussian Naive Bayes model through the predictive error. We evaluate our approaches on three real-life datasets that involve both long-term and short-term time dependence. Empirical studies demonstrate that our proposed techniques outperform the best-known competitors, which is a preferable choice for detecting anomalies.

In the era of data deluge, data analysis has become a key task for many industrial applications, e.g., master data management, and data integration. In particular, similarity join is an important primitive operator to support data analysis, which is to find similar pairs based on similarity functions and thresholds. In this article, we first propose a new similarity join operation called the dynamic skyline join without having to specify any similarity function or similarity threshold, which measures the similarity through multicriteria optimization. The dynamic skyline join operator makes the similarity join more flexible to support different criteria in multidimensional space. However, it is nontrivial to achieve dynamic skyline joins as both join operations and dynamic skyline queries are computationally complex in the increasing volume of real-world data. Therefore, we further propose Grid-SkyJoin, a framework to enable efficient parallel dynamic skyline joins on a shared-nothing cluster. Specifically, we use a grid partitioning to facilitate the data filtering and grouping strategies to provide load balancing and reduce the number of replicas. We also propose a multilevel filtering scheme to prune away a large fraction of unpromising points that do not fit into join results without actual join operations. Extensive experiments using benchmark datasets demonstrate that our filtering scheme can greatly reduce the number of data points to be joined, and our approach is about two times faster compared with the straightforward method in average.

The blooming of Internet of Things enables our modern buildings to become more and more smart, e.g., the intelligent LED lighting system could automatically adjust LEDs' dimming levels based on the ambient sunlight. However, this brings blurring effects that could significantly affect the effectiveness of visible light localization. In this paper, we propose a high-precision visible light localization system DIMLOC under dimmable LEDs in smart buildings. We first propose to use a novel imaging processing framework that consists of efficient techniques of the second-order polynomial fitting, histogram equalization, and Sobel filter to cope with the blurring effects. We then propose a novel algorithm based on vision analysis and scaling factor that only needs two LEDs for positioning. We prototype our DIMLOC on commercial off-the-shelf devices, and extensive experiments demonstrate that DIMLOC could achieve centimeter precision, e.g., 4.5 cm. Overall, this paper could further broaden application scenarios of visible light localization in the era of smart buildings.

Data aggregation is a fundamental operation in Internet of Things (IoT) applications, e.g., distributed Internet-based industrial control and computing systems. As IoT devices are increasingly connected to the system via resource-constrained wireless communication links, it is critical to perform communication-efficient data aggregation to answer complex queries (e.g., skyline queries and equality joins) from IoT applications. In this paper, we investigate the problem of constructing an aggregation tree (AT) for complex queries with the minimum communication cost. As complex queries have a dynamic size of intermediate results, existing Steiner tree-based approaches for traditional query operators, e.g., MIN and top-${k}$ , cannot be directly applied. We first formalize the aggregation gain by jointly considering the data pruning power (the size of data points that can be pruned during the aggregation for complex queries) and aggregation cost (the size of data points transmitted for the aggregation). By maximizing the aggregation gain, the data set that has a higher pruning power and a smaller size is selected and transferred for data aggregation at succeeding nodes. We then propose to construct the AT by connecting a set of aggregation operations with maximum aggregation gain. Extensive evaluation shows that our proposed framework achieves the promising results.

In-memory injection attacks are extremely challenging to reverse engineer because they operate stealthily without leaving artifacts in the system or in any easily observable events from outside of a virtual machine. Because these attacks perform their actions in memory only, current malware analysis solutions cannot expose their behavior. This paper introduces FAROS^1 a reverse engineering tool for Windows malware analysis based on dynamic information flow tracking (DIFT), which can flag stealthy in-memory-only malware injection attacks by leveraging the synergy of: (i) whole-system taint analysis; (ii) per security policy-based handling of the challenge of indirect flows via the application of tags of different types, and (iii) the use of tags with fine-grained provenance information. We evaluated FAROS with six advanced in-memory-injecting malware and it flagged the attacks for all samples. We also analyzed FAROS' false positive rate with 90 non-injecting malware samples and 14 benign software from various categories. FAROS presented a very low false positive rate of 2%, which shows its potential towards practical solutions against advanced in-memory-only anti-reverse-engineering attacks.

The goal of cyber attack investigation is to fully reconstruct the details of an attack, so we can trace back to its origin, and recover the system from the damage caused by the attack. However, it is often difficult and requires tremendous manual efforts because attack events occurred days or even weeks before the investigation and detailed information we need is not available anymore. Consequently, forensic logging is significantly important for cyber attack investigation. In this paper, we present DroidForensics, a multi-layer forensic logging technique for Android. Our goal is to provide the user with detailed information about attack behaviors that can enable accurate post-mortem investigation of Android attacks. DroidForensics consists of three logging modules. API logger captures Android API calls that contain high-level semantics of an application. Binder logger records interactions between applications to identify causal relations between processes, and system call logger efficiently monitors low-level system events. We also provide the user interface that the user can compose SQL-like queries to inspect an attack. Our experiments show that Droid Forensics has low runtime overhead (2.9% on average) and low space overhead (105 ~ 169 MByte during 24 hours) on real Android devices. It is effective in the reconstruction of realworld Android attacks we have studied.

Broadband mobile networks utilize a radio resource control (RRC) state machine to allocate scarce radio resources. Current implementations introduce high latencies and cross-layer degradation. Recently, the RRC enhancements, continuous packet connectivity (CPC), and the enhanced forward access channel (Enhanced FACH), have emerged in UMTS. We measure the availability and performance of these enhancements on a network serving a market with a population in the millions. We demonstrate that these enhancements offer significant reductions in latency, mobile device energy consumption, and improved end user experience. We develop new over-the-air measurements that resolve existing limitations in measuring RRC parameters. We find CPC provides significant benefits with minimal resource costs, prompting us to rethink past optimization strategies. We examine the crosslayer performance of CPC and Enhanced FACH, concluding that CPC provides reductions in mobile device energy consumption for many applications. While the performance increase of HS-FACH is substantial, cross-layer performance is limited by the legacy uplink random access channel (RACH), and we conclude full support of Enhanced FACH is necessary to benefit most applications. Given that UMTS growth will exceed LTE for several more years and the greater worldwide deployment of UMTS, our quantitative results should be of great interest to network operators adding capacity to these networks. Finally, these results provide new insights for application developers wishing to optimize performance with these RRC enhancements.

Visible Light Communication (VLC) is a promising technique for high-speed, low-cost wireless services with the rapid development and wide deployment of Light-Emitting Diodes (LEDs). However, the broadcast nature of the VLC makes eavesdroppers easily intercept the light communication in various settings, e.g., offices, conference rooms and airport lobbies. Although previous work put forward physical layer mechanisms to improve VLC security, they ignored the impact of light reflection and channel correlation on VLC security, which is not practical for the indoor environment. In this paper, we propose a new eavesdropping-resilient framework to defend against eavesdropping attacks under both Single Input Single Output (SISO) and Multiple Input Single Output (MISO) models. We propose a random time reversal scheme, that makes transmitted signal automatically focus on legitimate receivers while interfering the eavesdropper's channel with VLC's multipath redundancy, time reversal, and random choice technique. We analyze the impact of channel correlation on VLC security under the MISO model and find that the system secrecy capacity decreases. Therefore, we propose to use Karhunen-LoEve (KL) transform to improve the system secrecy capacity. Finally, we conduct extensive simulations to show that, our framework can make the eavesdropper's Bit Error Rate (BER) be above 0.0038, which is the threshold of Forward Error Correction (FEC), even when he tries to attack with Constant Modulus Algorithm (CMA). Furthermore, the system secrecy capacity is improved up to 3 Bit/Sec/Hz.

The use of secure HTTP calls is a first and critical step toward securing the Android application data when the app interacts with the Internet. However, one of the major causes for the unencrypted communication is app developer's errors or ignorance. Could the paradigm of literally repetitive and ineffective emphasis shift towards emphasis as a mechanism? This paper introduces emphaSSL, a simple, practical and readily-deployable way to harden networking security in Android applications. Our emphaSSL could guide app developer's security development decisions via real-time feedback, informative warnings and suggestions. At its core of emphaSSL, we use a set of rigorous security rules, which are obtained through an in-depth SSL/TLS security analysis based on security requirements engineering techniques. We implement emphaSSL via the PMD and evaluate it against 75 open- source Android applications. Our results show that emphaSSL is effective at detecting security violations in HTTPS calls with a very low false positive rate, around 2%. Furthermore, we identified 164 substantial SSL mistakes in these testing apps, 40% of which are potentially vulnerable to man-in-the-middle attacks. In each of these instances, the vulnerabilities could be quickly resolved with the assistance of our highlighting messages in emphaSSL. Upon notifying developers of our findings in their applications, we received positive responses and interest in this approach.

Fiber-Wireless (FiWi) access network, which could provide an anytime-anywhere access for end users with high bandwidth capacity and long distance, is facing the challenge of resource allocation and optimization due to the complexity and diversity of traffic demands. Though network virtualization becomes a promising solution, which allows heterogeneous virtual networks coexisting on the shared substrate network, previous works ignored both varied requirements of Quality of Service (QoS) satisfaction of virtual networks and the flexibility of reconfiguring the resource of substrate FiWi access network. In this paper, we propose a new Virtual Network Embedding (VNE) framework based on QoS satisfaction and network reconfiguration. By equipping each virtual network with a specific QoS satisfaction requirement, the characteristics of virtual network demands are formulated from a more practical point of view. Moreover, the adaptive bandwidth allocation of substrate network and virtual network reconfiguration are exploited to maximize the InP revenue. Simulation results demonstrate that our proposed VNE algorithm outperforms previous approaches with multifold increment of InP revenue.

We propose a new security paradigm that makes cross-layer personalization a premier component in the design of security solutions for computer infrastructure and situational awareness. This paradigm is based on the observation that computer systems have a personalized usage profile that depends on the user and his activities. Further, it spans the various layers of abstraction that make up a computer system, as if the user embedded his own DNA into the computer system. To realize such a paradigm, we discuss the design of a comprehensive and cross-layer profiling approach, which can be adopted to boost the effectiveness of various security solutions, e.g., malware detection, insider attacker prevention and continuous authentication. The current state-of-the-art in computer infrastructure defense solutions focuses on one layer of operation with deployments coming in a "one size fits all" format, without taking into account the unique way people use their computers. The key novelty of our proposal is the cross-layer personalization, where we derive the distinguishable behaviors from the intelligence of three layers of abstraction. First, we combine intelligence from: a) the user layer, (e.g., mouse click patterns); b) the operating system layer; c) the network layer. Second, we develop cross-layer personalized profiles for system usage. We will limit our scope to companies and organizations, where computers are used in a more routine and one-on-one style, before we expand our research to personally owned computers. Our preliminary results show that just the time accesses in user web logs are already sufficient to distinguish users from each other,with users of the same demographics showing similarities in their profiles. Our goal is to challenge today's paradigm for anomaly detection that seems to follow a monoculture and treat each layer in isolation. We also discuss deployment, performance overhead, and privacy issues raised by our paradigm.

Smartphone privacy and security work has focused mostly on malicious apps. We take a different angle by questioning whether good apps suffer from a lack of judgment and interact with "bad" websites. We use the term bad websites to refer to entities that engage in dangerous or annoying activities that range from distributing malware, to phishing and overly aggressive ad spamming. The focus of our work is this relatively neglected aspect of security: "Whom does an app talk to?" In this paper, we design and implement AURA, a framework for identifying the hosts that an app talks to and evaluating the risks this communication entails. AURA makes use of both static and dynamic analysis. We studied 13,500 popular free Android apps that connect to 254,022 URLs and 1,260 malicious Android apps that connect to 19,510 URLs. Our main contribution is showing that good apps pose security risks as they contact at least one website that: (a) distributes malware (8.8% of apps), (b) are in a blacklist (15% of apps) based on the classification by VirusTotal and Web of Trust. Our work can raise awareness that even good apps need to be carefully evaluated, especially as people become more concerned about smartphone security and privacy.

In current data centers, cloud providers (CPs) need to make decisions on virtual machine (VM) deployment requests that arrive dynamically over time under the constraint of limited physical resources to host VMs. Existing schemes to solve this problem either do not consider the dynamic characteristic of requests when solving the VM provisioning problem or separate admission control and VM provisioning into two independent problems. Both approaches can result in substantially sub-optimal revenue outcomes. In this paper, we propose a joint admission control and VM provisioning scheme based on a Markov Decision Process (MDP) framework to optimize the VM admission and placement decision in an integrated manner to maximize the CP's operating revenue. Because of the large number of physical servers in data centers, the MDP problem is computationally intractable to standard dynamic programming techniques. We propose a simulation-based approximate dynamic programming (ADP) algorithm to effectively solve the problem. Simulation results show that the new joint admission control and VM provisioning algorithm can substantially increase the revenue generated for CPs.

The Bring-Your-Own-Handheld-device (BYOH) phenomenon continues to make inroads as more people bring their own handheld devices to work or school. While convenient to device owners, this trend presents novel management challenges to network administrators. Prior efforts only focused on studying either the comparative characterization of aggregate network traffic between BYOHs and non-BYOHs or network performance issues, such as TCP and download times or mobility issues. We identify one critical question that network administrators need to answer: how do these BYOHs behave individually? In response, we design and deploy Brofiler, a behavior-aware profiling framework that improves visibility into the management of BYOHs. The contributions of our work are two-fold. First, we present Brofiler, a time-aware device-centric approach for grouping devices into intuitive behavioral groups. Second, we conduct an extensive study of BYOHs using our approach with real data collected over a year, and highlight several novel insights on the behavior of BYOHs. These observations underscore the importance of that BYOHs need to be managed explicitly as they behave in unique and unexpected ways.

Broadband mobile networks utilize a radio resource control (RRC) state machine to allocate scarce radio resources. Current implementations introduce high latencies and cross-layer degradation. Recently, the RRC enhancements, continuous packet connectivity (CPC) and the enhanced forward access channel (Enhanced FACH), have emerged in UMTS. We study the availability and performance of these enhancements on a network serving a market with a population in the millions. Our experience in the wild shows these enhancements offer significant reductions in latency, mobile device energy consumption, and improved end user experience. We develop new over-the-air measurements that resolve existing limitations in measuring RRC parameters. We find CPC provides significant benefits with minimal resource costs, prompting us to rethink past optimization strategies. We examine the cross-layer performance of CPC and Enhanced FACH, concluding that CPC provides reductions in mobile device energy consumption for many applications. While the performance increase of HS-FACH is substantial, cross-layer performance is limited by the legacy uplink random access channel (RACH), and we conclude full support of Enhanced FACH is necessary to benefit most applications. Given that UMTS growth will exceed LTE for several more years and the greater worldwide deployment of UMTS, our quantitative results should be of great interest to network operators adding capacity to these networks. Finally, these results provide new insights for application developers wishing to optimize performance with these RRC enhancements.

In this paper, we study the intertwined propagation of two competing "memes" (or data, rumors, etc.) in a composite network. Within the constraints of this scenario, we ask two key questions: (a) which meme will prevail? and (b) can one influence the outcome of the propagations? Our model is underpinned by two key concepts, a structural graph model (composite network) and a viral propagation model (SI1I2S). Using this framework, we formulate a non-linear dynamic system and perform an eigenvalue analysis to identify the tipping point of the epidemic behavior. Based on insights gained from this analysis, we demonstrate an effective and accurate prediction method to determine viral dominance, which we call the EigenPredictor. Next, using a combination of synthetic and real composite networks, we evaluate the effectiveness of various viral suppression techniques by either a) concurrently suppressing both memes or b) unilaterally suppressing a single meme while leaving the other relatively unaffected.

Android uses a system of permissions to control how apps access sensitive devices and data stores. Unfortunately, we have little understanding of the evolution of Android permissions since their inception (2008). Is the permission model allowing the Android platform and apps to become more secure? In this paper, we present arguably the first long-term study that is centered around both permission evolution and usage, of the entire Android ecosystem (platform, third-party apps, and pre-installed apps). First, we study the Android platform to see how the set of permissions has evolved; we find that this set tends to grow, and the growth is not aimed towards providing finer-grained permissions but rather towards offering access to new hardware features; a particular concern is that the set of Dangerous permissions is increasing. Second, we study Android third-party and pre-installed apps to examine whether they follow the principle of least privilege. We find that this is not the case, as an increasing percentage of the popular apps we study are overprivileged. In addition, the apps tend to use more permissions over time. Third, we highlight some concerns with pre-installed apps, e.g., apps that vendors distribute with the phone; these apps have access to, and use, a larger set of higher-privileged permissions which pose security and privacy risks. At the risk of oversimplification, we state that the Android ecosystem is not becoming more secure from the user's point of view. Our study derives four recommendations for improving the Android security and suggests the need to revisit the practices and policies of the ecosystem.

Survivability is one of the key issues in Fiber-Wireless (FiWi) broadband access network since huge data loss could be caused by single segment failure. Previous schemes focus on protecting FiWi against single segment failure by deploying backup fibers. However, these schemes suffer from two key problems. First, they ignore optimizing the selection of backup ONUs, which determines the recovery delay of the traffic interrupted by failure. Second, they underutilize the residual capacity of segments, thus require high cost of backup fibers. In this paper, aiming to tackle the inefficiency of previous schemes, we propose an efficient protection scheme, called Optimizing Backup ONUs selection and backup Fibers deployment (OBOF), to enhance the survivability of FiWi against single segment failure. Extensive experimental results demonstrate that our OBOF scheme outperforms the previous schemes significantly, especially in the scenario of higher traffic demand.

The Android platform lacks tools for assessing and monitoring apps in a systematic way. This lack of tools is particularly problematic when combined with the open nature of Google Play, the main app distribution channel. As our key contribution, we design and implement ProfileDroid, a comprehensive, multi-layer system for monitoring and profiling apps. Our approach is arguably the first to profile apps at four layers: (a) static, or app specification, (b) user interaction, (c) operating system, and (d) network. We evaluate 27 free and paid Android apps and make several observations: (a) we identify discrepancies between the app specification and app execution, (b) free versions of apps could end up costing more than their paid counterparts, due to an order of magnitude increase in traffic, (c) most network traffic is not encrypted, (d) apps communicate with many more sources than users might expect---as many as 13, and (e) we find that 22 out of 27 apps communicate with Google during execution. ProfileDroid is the first step towards a systematic approach for (a) generating cost-effective but comprehensive app profiles, and (b) identifying inconsistencies and surprising behaviors.

Landmark-based routing (LR) provides a promising approach for scalable point-to-point routing in wireless sensor networks (WSNs). Though various approaches have been proposed for landmark-based routing, they either introduce significant computational complexity or are inefficient in realistic, dynamic environments. In this paper, we identify three design principles that could form the basis of efficiency: algorithmic simplicity, update efficiency, and application awareness. Motivated by these principles, we present XLR, a new, flexible and comprehensive framework that tackles the inefficiency of landmark-based routing. XLR consists of four components: Relay Selection (RS), Parametric P-Norm distance function (PPN), Efficient Update with Coordinate Difference (EUCD) and General Forwarding (GF). The key advantage of XLR is that any subset of XLR's components can be independently incorporated into most landmark-based routing protocols.We perform extensive simulations to demonstrate that: (i) RS, a simple method, yields good performance comparable with previous methods, (ii) PPN increases LR performance considerably, (iii) EUCD reduces coordinate update overhead by up to 39%, and (iv) our GF outperforms previous approaches that consider factors such as link quality, delay and power consumption independently.