The integration of Domain Knowledge Graphs (DKG) into Retrieval-Augmented Generation has emerged as a promising approach for constructing Specialized Large Language Models (spLLMs).On account of the scarcity of high-quality DKGs, existing approaches employ an evolutionary framework, wherein the DKG is continuously evolved alongside its utilization for enhancing the LLM. Yet, these methods face two key limitations: 1) heavy reliance on costly expert knowledge, and 2) neglect of the connection between the LLM's inherent knowledge and external expert knowledge. To address these issues, this paper introduces Epistemic Cognition-enhanced Specialized LLMs (EC-spLLM), a novel evolutionary framework that instills epistemic cognition into the LLM to systematically exploit both its internal knowledge and expert knowledge. At the core of EC-spLLM lies the Hallucination Detection-based Epistemic Cognition (HDEC) mechanism, which assesses the reliability of LLM-generated responses using the LLM's self cognition and hallucination detection. This assessment ability enables EC-spLLM to selectively adopt either the expert-provided golden answer or a reliable LLM-generated answer during DKG evolution, thereby reducing dependence on experts and bridging internal and external knowledge sources to enhance performance. We conducted extensive experiments on five datasets spanning five domain, e.g., emotional sociology, biology, ect. Results show that EC-spLLM reduces the usage of golden answers by an average of 67% while retaining 97.2% of the accuracy achieved by the SOTA method, and outperforms all other baselines.

Cross-view geo-localization (CVGL) establishes correspondences between ground-level and satellite images of the same geographic location, serving as a fundamental technology for smart city applications, including autonomous navigation, urban planning, and location-based services. Current CVGL approaches fall into two categories: feature-based methods achieve superior performance through 2D representation learning but lack interpretability. Spatial-based methods provide geometric understanding and interpretable matching but suffer from limited spatial modeling and weak cross-view alignment, leading to lower performance. We reformulate CVGL from a spatial perspective and propose an auxiliary task-enhanced network. The network captures spatial semantics and provides explicit alignment processes with visualizable results. We introduce an auxiliary spatial semantic alignment (SSA) task that learns spatial structure via vision foundation models (VFM) and BEV transformation to enhance the primary CVGL task. The primary task captures visual semantics, including texture and appearance. Within this unified framework with shared encoders, the primary task enriches the learned embeddings by fusing spatial structure with visual semantics, yielding spatially complete representations. Extensive experiments on three standard CVGL benchmarks demonstrate that our method significantly surpasses previous spatial-based approaches while maintaining competitive performance with state-of-the-art (SOTA) feature-based methods, achieving 98.48% R@1 on CVUSA and 71.05% R@1 on CVACT\_test. We provide comprehensive analyses through pixel-level activation maps and feature-space UMAP visualizations to validate both effectiveness and interpretability.

Web 3.0-enabled edge computing provides a promising foundation for decentralized application (DApp) deployment. However, optimizing execution modes for interdependent task patterns remains challenging due to high computational overhead and blockchain consensus latency. This paper proposes an elastic hybrid architecture for Web 3.0-Enabled Edge Computing Systems (EDGEWEB3.0) that integrates on-chain and off-chain execution. DApp contracts are modeled as Directed Acyclic Graphs (DAGs) and partitioned into dependent task patterns, enabling a structured decomposition of execution. We formulate a DApp task scheduling problem and introduce the Decision Transformer-based Dependent Scheduling (DTDS) algorithm. DTDS employs masked self-attention to capture DAG dependencies and incorporates regularization terms to enforce constraints on service delays, gas costs, and computing capacity. We implement a real-world Web 3.0 testbed based on Ethereum, Goerli, and zkSync. Experimental results demonstrate that DTDS consistently outperforms state-of-the-art baselines in gas costs and service delays, providing a scalable solution for DApp execution in environments.

In the evolving landscape of Web 3.0, deploying and scheduling decentralized applications (DApps) presents significant challenges due to the complexity of heterogeneous nodes, edges, and their intricate interactions. Traditional approaches, particularly graph-based reinforcement learning (RL) methods, often rely on homogeneous graphs, which fail to capture the diverse relationships inherent in heterogeneous Web 3.0 environments. This limitation results in inefficient resource allocation, suboptimal task scheduling, and unclear security requirements. To address these issues, this paper introduces the Heterogeneous Graph Deployment Scheduler (HGDS), a novel framework that leverages Heterogeneous Graph Neural Networks (HGNNs) to model users, edge servers, and DApp tasks within Web 3.0 environments, and incorporates RL to optimize DApp scheduling policies. HGDS captures heterogeneity in Web 3.0 environments by jointly modeling node–edge interactions and heterogeneous edge relationships, enabling the generation of dynamic, task-aware embeddings that integrate both node and edge features. RL is further employed to adaptively optimize scheduling and resource allocation based on real-time network feedback. Experiments on a Web 3.0 testbed show that HGDS outperforms baseline methods by 12.2% in reward, while reducing service delay and gas consumption.

The widespread adoption of encryption across web-based services is compelling both malicious attackers and network defenders to tailor their tool repositories to encrypted traffic. For various security applications in encrypted networks, the analysis of encrypted traffic lies as the fundamental basis. Due to the inherent concealment of content-related information in encrypted packets, the dynamics of encrypted traffic emerge as the discernible variable warranting comprehensive analysis. This paper explores inherent temporal correlations within the encrypted traffic and proposes a novel algorithm calledCriss-crossTrafficTransformer (CTT), tailored to address unique challenges in encrypted traffic analysis. CTT distinguishes itself by employing a specialized time series Transformer that innovatively utilizespatchingandcriss-cross attention module(CAM) to dissect and interpret encrypted traffic, with the “criss” part mining the long-/short-term temporal correlations across time, and the “cross” part capturing temporal correlations across multiple feature dimensions of encrypted traffic. CTT provides a unified framework capable of accommodating diverse analytical granularities, including packet-level, flow-level, and packet-to-flow level. Notably, CTT not only encompasses encrypted traffic classification but also extends to encrypted traffic forecasting, an area that remains largely underexplored in existing literature. We evaluate CTT in the context of fingerprinting attacks and malware detection over 5 real-world datasets against 13 benchmarks. The results indicate that CTT achieves up to 15.56% performance improvement over SOTA solutions for encrypted traffic classification. Particularly, CTT demonstrates over 92.5% forecasting accuracy, which is comparable to SOTA performances in the seen-and-classify scenario, suggesting its potential applicability to broader domains like social network behavioral analysis. Our code is available athttps://github.com/Amanda-HuaDing/Criss-cross_Traffic_Transformer.

Large language models (LLMs) have demonstrated exceptional performance across a wide variety of domains. Nonetheless, generalist LLMs continue to fall short in reasoning tasks necessitating specialized knowledge, e.g., emotional sociology and medicine. Prior investigations into specialized LLMs focused on domain-specific training, which entails substantial efforts in domain data acquisition and model parameter fine-tuning. To address these challenges, this paper proposes the Way-to-Specialist (WTS) framework, which synergizes retrieval-augmented generation with knowledge graphs (KGs) to enhance the specialized capability of LLMs in the absence of specialized training. In distinction to existing paradigms that merely utilize external knowledge from general KGs or static domain KGs to prompt LLM for enhanced domain-specific reasoning, WTS proposes an innovative ''LLM↻KG'' paradigm, which achieves bidirectional enhancement between specialized LLM and domain knowledge graph (DKG). The proposed paradigm encompasses two closely coupled components: the DKG-Augmented LLM and the LLM-Assisted DKG Evolution. The former retrieves question-relevant domain knowledge from DKG and uses it to prompt LLM to enhance the reasoning capability for domain-specific tasks; the latter leverages LLM to generate new domain knowledge from processed tasks and use it to evolve DKG. WTS closes the loop between DKG-Augmented LLM and LLM-Assisted DKG Evolution, enabling continuous improvement in the domain specialization as it progressively answers and learns from domain-specific questions. We validate the performance of WTS on 7 datasets (e.g., TweetQA, ChatDoctor5k) spanning 6 domains, e.g., emotional sociology, medical, ect. The experimental results show that WTS surpasses the previous SOTA in 5 specialized domains, and achieves a maximum performance improvement of 11.3%.

Industrial defect detection is crucial for the production process, which identifies potential faults and minimizes manufacturing losses. Manufacturing operations are often distributed across multiple production lines and sites, with diverse defect samples scattered across edge devices. This necessitates the development of robust techniques to construct comprehensive detection models through collaborative learning among edge nodes. Federated learning (FL) has emerged as a promising paradigm for training such models in industrial edge environments while preserving data privacy. However, the presence of biased production conditions across edge nodes introduces statistical heterogeneity, which poses a significant challenge to the practical application of FL and adversely impacts the performance of aggregated global models. In this paper, we address the issue of statistical heterogeneity by leveraging hard sample mining and pseudo data generation. We propose a novel framework, namely Federated Hard Example Mining and Generation (FedHEM), which collects class prototypes and hard features from local clients and generates corresponding pseudo samples to transfer those knowledge. FedHEM can cope with statistical heterogeneity and satisfy the need for generalization ability and learning of specific classes. The proposed framework is evaluated on benchmark datasets and industrial defect detection tasks. Experimental results demonstrate that FedHEM achieves over 10% accuracy improvement compared to state-of-the-art data-heterogeneous methods and exhibits robust adaptability to real-world industrial environments.

Preventive maintenance is crucial in the industrial production process, enabling the prediction of potential faults and minimizing manufacturing losses. The industrial system consists of distributed heterogeneous end devices, which pose significant challenges for establishing generalized and effective solutions for preventive maintenance. Federated learning (FL) has emerged as a promising paradigm for training comprehensive models in industrial edge environments and protecting data privacy. However, model heterogeneity among devices presents a significant challenge in the application of FL to real-world industrial scenarios, hindering the aggregation and sharing of knowledge among edge models. In this paper, we address the issue of model heterogeneity by leveraging the concept of multi-model ensemble. We propose a novel framework, namely Federated Learning with Diversified Model Ensemble (FedDME), which collects knowledge of heterogeneous models from edge devices and integrates them through knowledge distillation. FedDME can cope with model heterogeneity and save communication costs to improve FL efficiency. The proposed approach is implemented within defect detection tasks related to preventive maintenance. The results demonstrate that our method achieves over 10% accuracy improvement compared with the state-of-the-art solution and exhibits robust adaptability to real-world industrial environments.

The confluence of Internet-of-Things (IoT) and artificial intelligence has advanced smart home (SH) systems, enabling the provision of complex scene-aware services. Central to these services is the precise perception of the indoor environment. Indoor scenes present unique challenges due to diverse layouts, frequent object occlusions, and dynamic human activities, which hinder the comprehensive understanding by individual SH devices/sensors. Moreover, the diversity of sensors equipped by SH devices introduces the multimodal data issue, necessitating the reconciliation of discrepancies among various data modalities. This article presents multimodal collaborative perception (MMCP), a collaborative perception paradigm for SH systems with multimodal raw data. MMCP leverages the intermediate collaboration framework and tailors it to an edge-assisted SH system. It deploys dedicated encoders at SH devices to convert multimodal raw data to uniform intermediate features, which are then sent to an edge computing box for aggregation and perception. MMCP introduces a critical information identifier to selectively transmit informative parts within intermediate features, thereby mitigating the communication overhead for bandwidth-constrained SH devices. Moreover, MMCP designs collaborative infomax (CIM) to facilitate intermediate feature aggregation. CIM defines multiview mutual information (MVMI) to capture dependencies between the aggregated feature and individual intermediate features from multiple SH devices. It employs contrastive learning to estimate and maximize MVMI in an unsupervised manner, such that the aggregated feature can retain discriminative information from individual intermediate features. We evaluate MMCP in four real-world indoor scene datasets. Experimental results show that MMCP outperforms noncollaborative strategy by 18% in average precision (AP). Particularly, MMCP strikes a favorable balance between perception performance and communication overhead, compressing intermediate features to a ratio of 13% while maintaining higher AP compared to state-of-the-art methods.

Edge computing-enabled virtual reality (VR) is increasingly explored in smart healthcare systems due to its potential to deliver immersive, real-time medical services. However, ensuring ultra-low latency and interpretable decision-making in such systems remains a significant challenge. In this paper, we propose an explainable deep reinforcement learning (XDRL)-enhanced VR video streaming framework for immersive healthcare systems to provide smooth and reliable VR services. Specifically, we first model the VR content analysis process at the edge sides and formulate a joint caching, communication, and computing (3C) resource optimization problem to maximize the VR quality of service (QoS) and minimize service latency. To address this complex 3C resource scheduling decision problem, we propose HXRL, a novel implementation of explainable deep reinforcement learning (XDRL), specifically designed to optimize immersive healthcare VR services. Comprehensive experiments show that our HXRL method improves average tile bitrate by 18.3%, cache hit ratio by 24.7%, and reduces system latency by 32.5% compared to baselines on real-world VR healthcare datasets. Moreover, a class activation map (CAM)-based visual analysis is integrated to interpret the learned policies of our model, highlighting the spatial attention of decision-making and enhancing trustworthiness in medical contexts.

Smart homes depend on collaborative sequential prediction tasks to optimize energy consumption and appliance scheduling. Federated learning (FL) offers a promising approach by enabling decentralized model training to balance privacy and usability. Yet, standard FL techniques fail to effectively address data diversity and individual user preferences in smart home contexts. To address these issues, we propose P3FL: a Privacy-Preserving Personalized Federated Learning framework that integrates tailored model training and privacy enhancements for federated collaborative predictions and decision-making. Our framework introduces the Personalized Collaborative Decision-Making (PCDM) algorithm, which dynamically adapts to different household environments while ensuring privacy and personalization. P3FL combines a global model for knowledge aggregation with a personalized adaptation module to provide fine-tuned predictions based on user preferences, environmental factors, and device configurations. Theoretical convergence bounds analysis confirms the robustness and efficiency of PCDM under conditions of strong convexity, smoothness, and bounded variance. Extensive experiments on real-world smart home datasets demonstrate that P3FL outperforms state-of-the-art methods, with PCDM achieving a training accuracy of 92.14%. Our approach enhances operational efficiency and ensures personalized user satisfaction, privacy enhancement in smart homes.

Multi-agent perception (MAP) allows autonomous systems to understand complex environments by interpreting data from multiple sources. This paper investigates intermediate collaboration for MAP with a specific focus on exploring "good" properties of collaborative view (i.e., post-collaboration feature) and its underlying relationship to individual views (i.e., pre-collaboration features), which were treated as an opaque procedure by most existing works. We propose a novel framework named CMiMC (Contrastive Mutual Information Maximization for Collaborative Perception) for intermediate collaboration. The core philosophy of CMiMC is to preserve discriminative information of individual views in the collaborative view by maximizing mutual information between pre- and post-collaboration features while enhancing the efficacy of collaborative views by minimizing the loss function of downstream tasks. In particular, we define multi-view mutual information (MVMI) for intermediate collaboration that evaluates correlations between collaborative views and individual views on both global and local scales. We establish CMiMNet based on multi-view contrastive learning to realize estimation and maximization of MVMI, which assists the training of a collaborative encoder for voxel-level feature fusion. We evaluate CMiMC on V2X-Sim 1.0, and it improves the SOTA average precision by 3.08% and 4.44% at 0.5 and 0.7 IoU (Intersection-over-Union) thresholds, respectively. In addition, CMiMC can reduce communication volume to 1/32 while achieving performance comparable to SOTA. Code and Appendix are released at https://github.com/77SWF/CMiMC.

Web 3.0 and Edge computing are inherently compatible, making them an ideal combination for building a secure and efficient distributed service platform to support decentralized applications (DApps). This paper investigates an elastic hybrid computing architecture in Edge Web 3.0, allowing DApp tasks to be executed in a hybrid manner by integrating on-chain and off-chain execution. The principle is to transfer a portion of DApp to an off-chain execution environment, along with an appropriate result verification process, to enhance computing efficiency and reduce blockchain overhead. We formulate a DApp task scheduling problem that jointly optimizes the execution pattern and offloading decision of user tasks. A learning-based DApp task scheduling scheme is designed based on Proximal Policy Optimization (PPO) to minimize the gas cost and service delay of DApps. Particularly, we tailor PPO to handle the hard constraints of service delay, gas consumption, and computing capacity in Edge Web 3.0 by adding regularization terms in the learning objective function. We establish an Edge Web 3.0 testbed based on Goerli, ZkSync, and Ethereum to evaluate the proposed method. The experimental results show that our method outperforms state-of-the-art benchmarks.

Swarm Learning (SL) provides a secure distributed learning environment to edge computing (EC) networks by leveraging blockchain technology for certified participation, encrypted information transmission, and immutable data storage. However, vanilla SL faces scalability limitations due to system-wide model aggregation, which bottlenecks its communication and blockchain efficiency. This paper presents a novel framework called Peer-to-peer Swarm Learning Over consOrtium blOckchain$(\mathbf{PSLO}_{3})$to enhance the scalability of SL over EC networks. PSLO3proposes a peer-to-peer swarm learning (P2P-SL) mechanism that only requires local communications for P2P model aggregation, thereby reducing the communication overhead of vanilla SL for system-wide model aggregation. Furthermore, PSLO3delivers P2P-SL over consortium blockchain and strategically organizes the edge servers into subchains to minimize the overhead of P2P-SL over consortium blockchain. A subchain formation scheme is designed based on graph partitioning by jointly analyzing the topological property of EC networks, message-passing patterns of P2P-SL, and the overhead of cross-/intra-chain interactions. An evaluation environment is built based on the Wecross platform to evaluate the performance of PSLO3. Experimental results demonstrate that PSLO3provides a reduction of 81.1% in communication overhead and a reduction of 26.03% in blockchain cost compared to vanilla swarm learning while demonstrating comparable learning performances.

This paper proposes Meta Parallel Graph Neural Network (MPGNN) to establish a scalable Network Intrusion Detection System (NIDS) for large-scale Internet of Things (IoT) networks. MPGNN leverages a meta-learning framework to optimize the parallelism of GNN-based NIDS. The core of MPGNN is a coalition formation policy that generates meta-knowledge for partitioning a massive graph into multiple coalitions/subgraphs in a way that maximizes the performance and efficiency of parallel coalitional NIDSs. We propose an offline reinforcement learning algorithm, called Graph-Embedded Adversarially Trained Actor-Critic (G-ATAC), to learn a coalition formation policy that jointly optimizes intrusion detection accuracy, communication overheads, and computational complexities of coalitional NIDSs. In particular, G-ATAC learns to capture the temporal dependencies of network states and coalition formation decisions over offline data, eliminating the need for expensive online interactions with large IoT networks. Given generated coalitions, MPGNN employs E-GraphSAGE to establish coalitional NIDSs which then collaborate via ensemble prediction to accomplish intrusion detection for the entire network. We evaluate MPGNN on two real-world datasets. The experimental results demonstrate the superiority of our method with substantial improvements in F1 score, surpassing the state-of-the-art methods by 0.38 and 0.29 for the respective datasets. Compared to the centralized NIDS, MPGNN reduces the training time of NIDS by 41.63% and 22.11%, while maintaining an intrusion detection performance comparable to centralized NIDS.

Microservice (MS) structures a service application as a collection of independently deployable service modules, making it particularly suitable for delivering complex applications in distributed computing systems. This paper investigates MS architecture over Mobile Edge Computing (MEC) networks (hereafter referred to as EdgeMS) and studies an EdgeMS placement problem that aims to deploy MS modules over the MEC network in a manner that maximizes the reward of MS application providers. A novel algorithm called Dual-GNN Deep Deterministic Policy Gradient (DG-DDPG) is proposed to establish an intelligent EdgeMS placement policy for optimizing the location of MS modules and performing fractional computing resource allocation. DG-DDPG leverages the graph neural network (GNN) to comprehend the graph-structured information encapsulated in the MS application structure and MEC network. A dual-GNN core is constructed in DG-DDPG, one GNN for MS applications to distill knowledge from intricate connections between MS modules, and the other GNN for MEC networks to capture complicated interactions between edge sites when providing EdgeMS. DG-DDPG embeds the dual-GNN core in a DDPG-based reinforcement learning framework, which not only handles temporal dependencies between EdgeMS placement decisions for maximizing long-term reward but also supports continuous action space for enabling fractional resource allocation. In particular, the learning process of DG-DDPG is tailored to address hard constraints (i.e., computing capacity and MS application completeness) in the EdgeMS placement problem. We design constraint-based regularization terms and add them to the objective of DG-DDPG, which facilitates the identification of feasible placement decisions during learning. We carry out systematic experiments to evaluate the performance of DG-DDPG, and the results show that DG-DDPG outperforms state-of-the-art benchmarks in terms of reward, service delay and deployment cost.

Blockchain has revolutionized a variety of fields by providing decentralization, immutability, transparency, and auditability. This paper designs Blockchained Edge Resource Auction (BERA) for edge computing systems to allocate computing resources to application service providers (ASP) in a secure manner. BERA comprises two key components: Blockchain-based Sealed-Bid Auction (BSBA) and Graph Neural Network (GNN)-based Fraud Detection (GFD). BSBA designs smart contracts to realize sealed-bid auctions overhead blockchain. It incorporates the homomorphic commitment technique to guarantee the transactional privacy of ASPs’ bidding information and performs interval membership zero-knowledge proof to verify the legitimacy of auction results. While the privacy-preserving property of BSBA is desirable, the veiled bidding information tends to breed fraudulent behaviors. Therefore, GFD is further proposed to identify abnormal auction behaviors in BSBA without revealing bidding information of ASPs. GFD converts the blockchain data of BSBA to an auction behavioral graph of ASPs, and uses GNN to discover stealth frauds based on interactive patterns. In addition, we design a subgraph extraction scheme for GFD to improve its scalability. We implement BERA on a private Ethereum blockchain and successfully realize edge resource auctions. We simulate several types of auction frauds and identify them with GFD. The experimental results show that our method outperforms other benchmarks.

With increasingly strict data privacy regulations, federated learning (FL) has become one of the most often heard machine learning techniques due to its privacy-preserving trait. To efficiently implement the FL intelligence, researchers recently resort to a newly emerged computing paradigm, mobile-edge computing (MEC), and bring about a burst of works. However, most existing works neglect practical issues in MEC systems, e.g., device heterogeneity, unstable channel conditions, and unknown user mobility. Any of them, if not handled properly, can cause fatal failures to FL. This article proposed a novel FL framework, called multicore FL (MC-FL), to help FL intelligence land successfully on realistic MEC systems. A distinct feature of MC-FL is maintaining and training multiple global models (GMs) that exhibit different tradeoffs between learning performances and computational complexity. While this modification seems simple, it can effectively handle the device heterogeneity and device status variations, and improve the compatibility and robustness of FL. Furthermore, MC-FL employs a partial client participation scheme that allows participating clients to vary across time. This enables MC-FL to function under uncertain mobile environments. We rigorously prove the convergence of the designed MC-FL framework. In particular, we propose an online client scheduling scheme for MC-FL to judiciously schedule clients for training multiple GMs in a manner that minimizes the completion time of MC-FL. We also provide a service provisioning scenario with MC-FL to show how service subscribers could benefit from multiple GMs and improve their Quality of Experience (QoE). We evaluate our method on real-world data sets, and the results show that MC-FL outperforms state-of-the-art benchmarks.

The rapid uptake of intelligent applications is pushing deep learning (DL) capabilities to mobile devices. However, the heterogeneities in device capacity, DNN performances, and user preferences make it challenging to provide satisfactory Quality of Experience (QoE) to mobile users. This paper studies automated customization for DL inference on mobile devices (termed as on-device inference), and our goal is to enhance user QoE by configuring the on-device inference with an appropriate DNN for users under different usage scenarios. The core of our method is a DNN selection module that learns user QoE patterns on-the-fly and identifies the best-fit DNN for on-device inference with the learned knowledge. It leverages an online learning algorithm,NeuralUCB, that has excellent generalization ability for handling various user QoE patterns. We also embed the knowledge transfer technique in NeuralUCB to expedite the learning process. However, NeuralUCB frequently solicits QoE ratings from users, which incurs non-negligible inconvenience. To address this problem, we design feedback solicitation schemes to reduce the number of QoE solicitations while maintaining the learning efficiency of NeuralUCB. A pragmatic problem,aggregated QoE, is further investigated to improve the practicality of our framework. We conduct experiments on both synthetic and real-world data. The results indicate that our method efficiently learns the user QoE pattern with few solicitations and provides drastic QoE enhancement for mobile devices.

Advances in deep learning (DL) have triggered an explosion of mobile intelligence, posing a soaring demand for computing resources that cannot be satisfied by mobile devices. In this article, we employedge computingto deliver better DL inference services to end users. The key is to leverage deep neural network (DNN) ensemble techniques that provide state-of-the-art performance for many machine learning applications in terms of inference accuracy and robustness. Compared to end devices, the edge computing platform is endowed with more powerful computing resources, making it feasible to implement DNN ensembles for DL inferences. However, due to the constrained computing capacity of edge servers and the possible service response deadline, an edge server can only use a limited number of DNNs to construct DNN ensembles. This poses a unique problem, namely, DNN ensemble selection, for identifying the best-fit DNN ensembles. We propose a novel algorithm called automated DNN ensemble selection (AES) algorithm to solve this problem. Because DNNs exhibit performance variations over different distributions of input data, AES adaptively determines a DNN ensemble according to the features of admitted inference tasks. AES is an online learning algorithm that learns DNNs’ in-use performance over time. An ensemble selection rule is further designed as a subroutine of AES to recruit members into the DNN ensemble based on the accuracy and diversity of DNNs. In particular, we theoretically prove that AES can achieve asymptotic optimality. We carry out experiments on real-world data sets. The results show that using the DNN ensemble technique on edge computing platforms dramatically improves the DL inference quality, and AES outperforms other benchmark schemes.

The momentous enabling of deep learning (DL)-powered mobile application is posing a soaring demand for computing resources that can hardly be satisfied by mobile devices. In this paper, we employ Edge Computing to deliver DL inference services to mobile users, where Deep Neural Networks (DNNs) are configured on edge servers, processing inference tasks received from mobile devices. A novel method called Adaptive DNN Ensemble (ADE) is proposed to enhance the performance of DL inference services. The core of ADE is the DNN ensemble technique which improves the stability and accuracy of DL inference. Due to the limited computing resources and service response deadline, ADE needs to judiciously determine DNNs to be included in the DNN ensemble, which poses a unique DNN ensemble selection problem. In addition, because DNNs exhibit performance variations for tasks with different features, DNN ensemble selection also aims to reconFigure DNN ensembles according to the feature of admitted tasks. We design an online learning algorithm, Contextual Combinatorial Multi-Armed Bandit (CC-MAB), to learn the DNN performance for tasks with different features. We rigorously prove that the proposed online learning algorithm is able to achieve asymptotic optimality. Experiments are carried out on an edge computing testbed to evaluate our method. Various implementation concerns, including memory usage, time complexity, and DNN switching cost, are considered. The results show that ADE outperforms other benchmarks in terms of inference accuracy and can provide real-time responses.

Mobile Edge Computing (MEC) is delivering a rich portfolio of computation services to enable ultra-low latency and location-awareness for emerging mobile applications. However, the vulnerability of this new paradigm to potential security and privacy issues prevents mobile users from fully embracing its advantage. While various defensive strategies have been proposed to secure the connection between the end devices and edge servers, an equally important issue, the server-side risk is still under-investigated for most edge computing systems. To handle these server-side risks, a Risk-aware Computation Offloading (RCO) policy is proposed to distribute computation tasks safely among geographically distributed edge sites under server-side attacks. RCO takes into account the strategic behaviors of the potential attackers in the edge system and finds an appropriate balance between risk management and service delay reduction. The Bayesian Stackelberg game is employed to formulate the RCO problem, which describes an appropriate relation between the edge system (as a defender) and the attacker. In particular, the Bayesian Stackelberg game captures the uncertainty of attacker’s behavior and enables RCO to work even when the edge system does not know precisely the attacker that it is playing against. To facilitate the derivation of Stackelberg equilibria, two pruning rules, Heuristic Pruning (HP) and Branch-and-Bound (BaB), are proposed. HP prunes by analyzing the user demand distribution and attacker types, and BaB prunes by obtaining the tight upper/lower bound of edge system utility with the assist of disjunctive programming and Bender’s cut.