Large Language Models (LLMs) achieve strong performance through extended inference-time deliberation, yet how their reasoning failures arise remains poorly understood.By analyzing model-generated reasoning trajectories, we find that errors are not uniformly distributed but often originate from a small number of early transition points, after which reasoning remains locally coherent but globally incorrect.These transitions coincide with localized spikes in token-level entropy, and alternative continuations from the same intermediate state can still lead to correct solutions.Based on these observations, we introduce GUARD 1 , a targeted inference-time framework that probes and redirects critical transitions using uncertainty signals.Empirical evaluations across multiple benchmarks confirm that interventions guided by these failure dynamics lead to more reliable reasoning outcomes.Our findings highlight the importance of understanding when and how reasoning first deviates, complementing existing approaches that focus on scaling inference-time computation.

Towards Knowledge Graph (KG) completion, probabilistic logic reasoning approaches enable effective rule mining but incur high computational cost, while embedding-based methods offer high efficiency but confront limited semantic understanding. Neuro-symbolic approaches combine both by employing embeddings to approximate probabilistic distributions, yet face challenges in weight optimization and hurdles in scaling up from large probability graphs. To address these issues, we propose DPLogic, a differentiable probabilistic logic reasoning framework for KG completion. Initially, we construct a Markov logic network by selecting crucial formulas and constraining groundings to relevant subgraphs, effectively boosting the scalability of the framework. Subsequently, we represent formula weights through relation-specific embeddings by introducing neural logical operators, creating a differentiable pathway for end-to-end optimization. Finally, we obtain the distribution of unobserved KG triplets by facilitating the joint optimization of embedding-based and probabilistic distributions through an EM algorithm. Empirical findings on standardized datasets illustrate that our proposed DPLogic consistently surpasses state-of-the-art methodologies in terms of both efficacy and efficiency.

Multivariate time series (MTS) forecasting is crucial for predicting the future states of complexly coupled variables based on historical observations. To effectively capture the intricate interdependencies within MTS, graph-based methods have emerged as powerful tools. However, existing graph construction methods often produce structures that fail to preserve key temporal and cross-variable dependencies, introducing redundant or irrelevant connections. To address these challenges, we propose a structural entropy-based approach for MTS forecasting. The approach optimizes graph structures by reducing structural redundancy, thereby improving forecasting accuracy. Initially, we represent the temporal dependences by constructing an encoding tree incrementally. Through hierarchical organization of time steps, the temporal evolution is adaptively captured. Subsequently, we give the community-aware representation by building an encoding tree over the variables in MTS, extracting homogeneous communities from the tree structure while integrating community influence to better capture inter-variable dependencies. Finally,we present a training algorithm designed to generate accurate predictions for MTS, accompanied by a unified loss function that integrates forecasting inaccuracies with variations in structural entropy. Empirical findings on real-world datasets substantiate that our approach outperforms state-of-the-art models in capturing dependencies and enhancing forecasting precision.

Federated learning (FL) facilitates collaborative global model training without compromising data privacy. However, data distribution variations among clients inevitably introduce bias in global updates, impacting model fairness and performance. Existing methods assign client aggregation weights simply based on dataset size proportions or rely on substantial assumptions about specific global data distributions such as uniform label distributions. These approaches inadequately capture the intrinsic impact of Non-IID data characteristics on model divergence. To address these deficiencies, we propose a novel adaptive weight allocation algorithm, FedFMD, leveraging Mahalanobis distance, integrating Task Arithmetic, to dynamically assign weights based on client contributions. FedFMD explicitly models task-centric deviations caused by data heterogeneity without requiring raw data access or prior distribution assumptions. Besides, FedFMD enhances aggregation weights computation through time-decay adjustments, guided by historical client performance trends, optimizing both fairness and utility. Extensive evaluations against six state-of-the-art (SOTA) algorithms and two distance metrics across three datasets demonstrate the superior performance of FedFMD in fairness and utility.

The computational inefficiency of spiking neural networks (SNNs) is primarily due to the sequential updates of membrane potential, which becomes more pronounced during extended encoding periods compared to artificial neural networks (ANNs). This highlights the need to parallelize SNN computations effectively to leverage available hardware parallelism. To address this, we propose Membrane Potential Estimation Parallel Spiking Neurons (MPE-PSN), a parallel computation method for spiking neurons that enhances computational efficiency by enabling parallel processing while preserving the intrinsic dynamic characteristics of SNNs. Our approach exhibits promise for enhancing computational efficiency, particularly under conditions of elevated neuron density. Empirical experiments demonstrate that our method achieves state-of-the-art (SOTA) accuracy and efficiency on neuromorphic datasets. Codes are available at https://github.com/chrazqee/MPE-PSN.

Multivariate time series (MTS) forecasting endeavors to anticipate the forthcoming sequence of interdependent variables through the utilization of past observations. The prevailing methodologies, relying on deep neural networks, Transformer, or information bottleneck frameworks, persist in confronting challenges such as overlooking or inadequately capturing the inter / intra-series correlations evident in practical MTS datasets. In response to these challenges, we introduce a conditional information bottleneck-based strategy for MTS forecasting, grounded in information theory. Initially, we establish a conditional information bottleneck principle to capture the inter-series correlations via conditioning on non-target variables. Subsequently, a conditional mutual information-based technique is introduced to extract intra-series correlations by conditioning historical data, ensuring temporal consistency within each variable. Lastly, we devise a unified optimization objective and propose a training algorithm to collectively capture inter / intra-series correlations. Empirical investigations on authentic datasets underscore the superiority of our proposed approach over other cutting-edge competitors. Our code is available at https: //github.com/Xinhui-Lee/CIB-MTSF.

Graph contrastive learning (GCL) has demonstrated remarkable success in training graph neural networks (GNNs) by distinguishing positive and negative node pairs without human labeling. However, existing GCL methods often suffer from two limitations: the repetitive message-passing mechanism in GNNs and the quadratic computational complexity of exhaustive node pair sampling in loss function. To address these issues, we propose an efficient and effective GCL framework that leverages community structure rather than relying on the intricate node-to-node adjacency information. Inspired by the concept of sparse low-rank approximation of graph diffusion matrices, our model delivers node messages to the corresponding communities instead of individual neighbors. By exploiting community structures, our method significantly improves GCL efficiency by reducing the number of node pairs needed for contrastive loss calculation. Furthermore, we theoretically prove that our model effectively captures essential structure information for downstream tasks. Extensive experiments conducted on real-world datasets illustrate that our method not only achieves the state-of-the-art performance but also substantially reduces time and memory consumption compared with other GCL methods. Our code is available at [https://github.com/chenx-hi/IGCL-CS](https://github.com/chenx-hi/IGCL-CS).

As the general description of relationships between attributes, approximate functional dependencies (AFDs) almost hold for a given dataset with a few violations. Most of existing methods for AFD discover are insufficient to balance the efficiency and accuracy due to the massive search space and permission of violations. To address these issues, we propose an efficient method of probabilistic semantics guided discovery of AFDs based on Bayesian network (BN). Firstly, we learn a BN structure and conduct conditional independence tests on the learned structure rather than the entire search space, such that candidate AFDs could be obtained. Secondly, we fulfill search space reduction and structure pruning by making use of probabilistic semantics of graphical models in terms of BN. Consequently, we provide a branch-and-bound algorithm to discover the AFDs with the highest smoothed mutual information scores. Experimental results illustrate that our proposed method is more effective and efficient than the comparison methods. Our code is available at [https://github.com/DKE-Code/BNAFD](https://github.com/DKE-Code/BNAFD).

Federated learning (FL) is a widely used distributed learning framework. However, constrained wireless environment and intrinsically heterogeneous data across devices can hinder the FL framework being practical. In this paper, we propose a communication-efficient FL framework that helps boost the training process by considering the transmission power of each device and the local models' personalized training. In each round of training, we select the participating devices that can minimize the upper bound of the convergence rate plus the corresponding communication overhead while subjecting to the transmit power constraint. Besides, each device update a personalized and sparse model that only consumes limited computation resources. We validate our proposed FL framework on various dataset, and experiment results show that our framework speeds up the training process by taking$\sim$40% less time than the existing frameworks. Also, the communication time can be significantly decreased by employing our framework, e.g., we achieve as high as a 42.7% increase in test accuracy and save up to 74.3$\%$in the communication cost compared with FedAvg.

Computation offloading is indispensable for mobile edge computing (MEC). It uses edge resources to enable intensive computations and save energy for resource-constrained devices. Existing works generally impose strong assumptions on radio channels and network queue sizes. However, practical MEC systems are subject to various uncertainties rendering these assumptions impractical. In this paper, we investigate the energy-efficient computation offloading problem by relaxing those common assumptions and considering intrinsic uncertainties in the network. Specifically, we minimize the worst-case expected energy consumption of a local device when executing a time-critical application modeled as a directed acyclic graph. We employ the extreme value theory to bound the occurrence probability of uncertain events. To solve the formulated problem, we develop an$\epsilon $-bounded approximation algorithm based on column generation. The proposed algorithm can efficiently identify a feasible solution that is less than$(1+\epsilon)$of the optimal one. We implement the proposed scheme on an Android smartphone and conduct extensive experiments using a real-world application. Experiment results corroborate that it will lead to lower energy consumption for the client device by considering the intrinsic uncertainties during computation offloading. The proposed computation offloading scheme also significantly outperforms other schemes in terms of energy saving.

Internet connected mobile devices in 5G and beyond (simply 5G+) systems are penetrating all aspects of people's daily life, transforming the way we conduct business and live. However, this rising trend has also posed unprecedented traffic burden on existing telecommunication infrastructure including cellular systems, consistently causing network congestion. Although additional spectrum resources have been allocated, exponentially increasing traffic tends to always outpace the added capacity. In order to increase the data rate and reduce the latency, 5G+ systems have heavily relied on hyperdensification and higher frequency bands, resulting in dramatically increased interference temperature, and consequently significantly more weak signals (i.e., signals with low Signal-to-Noise-plus-Interference (SINR) ratio). With traditional detection mechanisms, a large number of weak signals will not be detected, and hence be wasted, leading to poor throughput in 5G+ systems.

The Internet of Things (IoT) integrates information collected from multisources and is able to support various intelligent smart city applications, such as industrial manufacturing, power systems, and mobile healthcare. In the big data era, multisourced data are collected on a daily basis, whereas a large part of the data may be irrelevant, redundant, noisy, or even malicious from a machine learning perspective. Feature selection has been a powerful data cleaning technique to reduce data redundancy and improve system performance in machine learning. Inspired by reinforcement learning that learns from its experience, in this article, we propose a novel efficient deep$Q$-network (DQN)-based feature selection method for multisourced data cleaning. In particular, we model the feature selection problem as a competition between an agent and the environment in dynamic states, which is solved by a DQN. Traditional DQN suffers from high computational complexity and requires a significant amount of time in order to converge in the training process. To tackle these challenges, we develop a space searching algorithm called SS to speed up the training process of the DQN agent. To validate the efficacy and efficiency of the proposed method, we conduct extensive experiments on various types of IoT data. Simulation results show that the proposed DQN-based feature selection algorithms achieve much better performance compared with state-of-the-art methods, and are robust under data poisoning attacks.

While traffic modeling and prediction are at the heart of providing high-quality telecommunication services in cellular networks and attract much attention, they have been approved as an extremely challenging task. Due to the diverse network demand of Internet-based apps, the cellular traffic from an individual user can have a wide dynamic range. Most existing methods, on the other hand, model traffic patterns as probabilistic distributions or stochastic processes and impose stringent assumptions over these models. Such assumptions may be beneficial at providing closed-form formula in evaluating prediction performances, but fall short for practice use. In this paper we propose STEP, aspatio-temporal fine-granular user trafficprediction mechanism for cellular networks. A deep graph convolution network, called GCGRN, is constructed. It is a novel combination of the graph convolution network (GCN) and gated recurrent units (GRU), which exploits graph neural network to learn an efficient spatio-temporal model from a user’s massive dataset for traffic prediction. The prototype of STEP has been implemented. Extensive experimental results demonstrate that our model outperforms the state-of-the-art time-series based approaches. Besides, STEP merely incurs mild energy consumption, communication overhead and system resource occupancy to mobile devices. Moreover, NS-3 based simulations validate the efficacy of STEP in reducing session dropping ratio in cellular networks.

Edge computing, an emerging computing paradigm pushing data computing and storing to network edges, enables many applications that require high computing complexity, scalability, and security. In the big data era, one of the most critical applications is multiparty learning or federated learning, which allows different parties to collaborate with each other to obtain better learning models without sharing their own data. However, there are several main concerns about the current multiparty learning systems. First, most existing systems are distributed and need a central server to coordinate the learning process. However, such a central server can easily become a single point of failure and may not be trustworthy. Second, although quite a few schemes have been proposed to study Byzantine attacks, a very common and challenging kind of attack in distributed systems, they generally consider the scenario of learning a global model. However, in fact, all parties in multiparty learning usually have their own local models. The learning methods and security issues, in this case, are not fully explored. In this article, we propose a novel blockchain-empowered decentralized secure multiparty learning system with heterogeneous local models called BEMA. Particularly, we consider two types of Byzantine attacks, and carefully design “off-chain sample mining” and “on-chain mining ” schemes to protect the security of the proposed system. We theoretically prove the system performance bound and resilience under Byzantine attacks. The simulation results show that the proposed system obtains comparable performance with that of conventional distributed systems, and bounded performance in the case of Byzantine attacks.

Community detection is an effective approach to unveil relationships among individuals in online social networks. In the literature, quite a few algorithms have been proposed to conduct community detection by exploiting the topology of social networks and the attributes of social actors. In practice, community detection is usually conducted by third parties, such as advertisement companies and hospitals, with access to social networks for different purposes, which can easily lead to a privacy breach. In this paper, we investigate community detection in social networks aiming to protect the privacy of both the network topology and the users' attributes. We show that with additional prior knowledge, community detection can be performed by querying the information of only a fraction of instead of the entire population. In particular, we first propose a new scheme called differentially private community detection (DPCD). DPCD detects communities in social networks via a probabilistic generative model, which can be decomposed into subproblems solved by individual users. The private social relationships and attributes of each user are protected by objective perturbation with differential privacy guarantees. Then, we propose a parsimonious node affiliation recovery (NAR) algorithm, which is also differentially private, to unveil the community affiliation information of the whole population based on that of the limited number of queried individuals by solving a sparse optimization problem. Through both theoretical analysis and experimental validation using synthetic and real-world social networks, we demonstrate that the proposed DPCD scheme detects social communities under the modest privacy budget. In addition, we show the effectiveness of NAR to perform community detection by querying a limited number of individuals in social networks.

The ever-growing proliferation of mobile devices equipped with accelerometers has provided new opportunities to capture the semantic meanings of human activities and improve user experience with behavior-based recommendations, which heavily rely on the accuracy of the recognition of daily human activities. Acceleration-based human activity recognition (HAR) is a challenging problem because each accelerometer records multi-dimensional signals in both spatial and temporal domains that have different attributes for representing different activities or even the same activity. Thus we cannot directly compare these signals with each other, because they are embedded in a non-metric space. In this paper, we present a Personalized Recurrent Neural Network (PerRNN) to dynamically segment and recognize the human activities based on accelerometer data. Enlightened by the idea of spatiotemporal predictive learning, the proposed architecture is capable of memorizing different acceleration signals' appearances and temporal variations in a unified memory pool. We evaluate the performance of the proposed framework on a commonly used dataset, WISDM. Experiment results show that compared with state-of-the-art schemes, our proposed PerRNN system recognizes 6 different human activities with the highest overall accuracy of 96.44%.

Studies find that deep learning models are vulnerable to deliberate adversarial manipulations by attackers. Adversarial training is an effective approach to address this problem. Previous works quantize the input sample space to find appropriate perturbations on the benign samples so as to generate adversarial samples for adversarial training. However, since only the input sample space is quantized with the perturbation space being still continuous, finding the optimal perturbation noise is still a non-convex and computationally expensive problem. Moreover, in this case, the found perturbation noise that will be used to generate an adversarial sample may be strong in the continuous search space, but may become weak after quantization in the input sample space. In this paper, we first develop an Iterative Quantized Local Search (IQLS) algorithm that finds strong perturbation noises by quantizing both the input space and perturbation space. Then, we theoretically analyze and prove the upper bound on the number of iterations needed for the IQLS algorithm, based on which we devise an efficient and effective Quantized Adversarial Training (QAT) scheme. Experiment results on six public datasets show that our proposed scheme outperforms state-of-the-art methods to defend against different adversarial attacks. Particularly, QAT improves the system performance by 14%, 11%, 16% on average on CIFAR-10, SVHN, and CIFAR-100 datasets respectively compared with the existing defense schemes, and reduces the computing time by about 60%.

In the big data era, distributed machine learning emerges as an important learning paradigm to mine large volumes of data by taking advantage of distributed computing resources. In this work, motivated by learning to learn, we propose a meta-learning approach to coordinate the learning process in the master-slave type of distributed systems. Specifically, we utilize a recurrent neural network (RNN) in the parameter server (the master) to learn to aggregate the gradients from the workers (the slaves). We design a coordinatewise preprocessing and postprocessing method to make the neural network based aggregator more robust. Besides, to address the fault tolerance, especially the Byzantine attack, in distributed machine learning systems, we propose an RNN aggregator with additional loss information (ARNN) to improve the system resilience. We conduct extensive experiments to demonstrate the effectiveness of the RNN aggregator, and also show that it can be easily generalized and achieve remarkable performance when transferred to other distributed systems. Moreover, under majoritarian Byzantine attacks, the ARNN aggregator outperforms the Krum, the state-of-art fault tolerance aggregation method, by 43.14%. In addition, our RNN aggregator enables the server to aggregate gradients from variant local models, which significantly improve the scalability of distributed learning.

Inference using deep neural networks may be outsourced to the cloud due to its high computational cost, which, however, raises security concerns. Particularly, the data involved in deep neural networks can be highly sensitive, such as in medical, financial, commercial applications, and hence should be kept private. Besides, the deep neural network models owned by research institutions or commercial companies are their valuable intellectual properties and can contain proprietary information, which should be protected as well. Moreover, an untrusted cloud service provider may return accurate and even erroneous computing results. To address the above issues, we propose a secure outsourcing framework for deep neural network inference called SecureNets, which can preserve both a user’s data privacy and his/her neural network model privacy, and also verify the computation results returned by the cloud. Specifically, we employ a secure matrix transformation scheme in SecureNets to avoid privacy leakage of the data and the model. Meanwhile, we propose a verification method that can efficiently verify the correctness of cloud computing results. Our simulation results on four- and five-layer deep neural networks demonstrate that SecureNets can reduce the processing runtime by up to $64%$. Compared with CryptoNets, one of the previous schemes, SecureNets can increase the throughput by $104.45%$ while reducing the data transmission size by $69.78%$ per instance.

Unsupervised anomaly detection on multidimensional time series data is a very important problem due to its wide applications in many systems such as cyber-physical systems, the Internet of Things. Some existing works use traditional variational autoencoder (VAE) for anomaly detection. They generally assume a single-modal Gaussian distribution as prior in the data generative procedure. However, because of the intrinsic multimodality in time series data, previous works cannot effectively learn the complex data distribution, and hence cannot make accurate detections. To tackle this challenge, in this paper, we propose a GRU-based Gaussian Mixture VAE system for anomaly detection, called GGM-VAE. In particular, Gated Recurrent Unit (GRU) cells are employed to discover the correlations among time sequences. Then we use Gaussian Mixture priors in the latent space to characterize multimodal data. The proposed detector reports an anomaly when the reconstruction probability is below a certain threshold. We conduct extensive simulations on real world datasets and find that our proposed scheme outperforms the state-of-the-art anomaly detection schemes and achieves up to 5.7% and 7.2% improvements in accuracy and F1 score, respectively, compared with existing methods.

The popularity of smart mobile devices has resulted in the surged growth of mobile data traffic, which makes current cellular communication systems overloaded. To accommodate the data, the current wireless communication system is evolving to a 5G wireless communication system that employs multiple technologies to boost its system capacity. We notice that non-line-of-sight (NLOS) transmission is ubiquitous in wireless communication systems, and is even more common in 5G wireless communication systems due to using millimeter-Wave (mmWave) communications. Previous works employ beamforming techniques to enhance NLOS transmission performance but suffer from the high cost for controlling antennas. In this paper, we propose a dynamic transmission power control scheme for improving NLOS transmission performance. Particularly, we explore the control of UE association with MBS/SBSs and power allocation to maximize UEs' sum-rate under the constraints of transmission power and UEs' quality of service (QoS). To solve this maximization problem, we propose a deep Q- network (DQN) scheme, in which we apply a convolutional neural network (CNN) to estimate the Q-function offline and conduct a deep Q-learning online to find the control strategy. We offer simulation results to show the efficacy of the proposed scheme.