Serverless computing is a popular cloud computing paradigm that enables developers to build applications at the function level, known as serverless applications. The Serverless Application Model (AWS SAM) is the most widely adopted configuration schema. However, misconfigurations pose a significant challenge due to the complexity of serverless configurations and the limitations of traditional data-driven techniques. Recent advancements in Large Language Models (LLMs), pre-trained on large-scale public data, offer promising potential for identifying and explaining misconfigurations. In this paper, we present SlsDetector, the first framework that harnesses the capabilities of LLMs to perform static misconfiguration detection in serverless applications. SlsDetector utilizes effective prompt engineering with zero-shot prompting to identify configuration issues. It designs multi-dimensional constraints aligned with serverless configuration characteristics and leverages the Chain of Thought technique to enhance LLM inferences, alongside generating structured responses. We evaluate SlsDetector on a curated dataset of 110 configuration files, which includes correct configurations, real-world misconfigurations, and intentionally injected errors. Our results show that SlsDetector, based on ChatGPT-4o (one of the most representative LLMs), achieves a precision of 72.88%, recall of 88.18%, and F1-score of 79.75%, outperforming state-of-the-art data-driven methods by 53.82, 17.40, and 49.72 percentage points, respectively. We further investigate the generalization capability of SlsDetector across recent LLMs, including Llama 3.1 (405B) Instruct Turbo, Gemini 1.5 Pro, and DeepSeek V3, with consistently high effectiveness.

AI is making mobile applications increasingly cooler, but also introduces serious privacy risks due to the extensive user data collection. Federated learning (FL), as a privacy-preserving machine learning paradigm, enables mobile devices to collaboratively learn a shared prediction model while keeping all training data on devices. However, a key obstacle towards practical cross-device FL training is the huge energy consumption, especially for lightweight mobile devices. Prior literature mostly optimizes the convergence speed and network communication cost. In this work, we first perform the experimental analysis of improving FL performance through low-precision training with energy-friendly Digital Signal Processor (DSP) on mobile devices. Then, we demonstrate that directly integrating the state-of-the-art INT8 (8-bit integer) training algorithm and classic FL protocols will significantly degrade the model accuracy. Finally, we propose a novel FL protocol, namely Q-FedUpdate, incorporates two critical techniques: error-compensated aggregation and pipelined batch quantization. The former can ensure the tiny model updates be accumulated and take effects, and the latter can improve the DSP cache hit rate to reduce the context switching. Extensive experiments show that, Q-FedUpdate can effectively reduce the on-device energy consumption by 21 & times;, and accelerate the FL convergence by 6.1 & times; with only 2% accuracy loss.

The proliferation of intelligent consumer electronics necessitates a service architecture capable of delivering personalized services while protecting privacy. Federated Learning (FL), as a privacy-preserving paradigm, enables multiple consumer electronic devices to collaboratively train a model without exposing sensitive user data. However, FL performance (i.e., convergence performance, global generalization, and local personalization) is severely compromised by heterogeneous data distribution across consumer electronic devices, which is prevalent in most real-world scenarios. Prompt learning, as an efficient approach, is introduced to improve either global generalization or local personalization. They typically struggle to balance both directions, as optimizing one often leads to failure in another. To address this problem, we design a novel FL framework named PGA-FL, which fuses local and global prompts by Genetic Algorithm (GA) to update the local prompts on the clients and aggregates the uploaded local prompts based on their semantic similarity to update the global prompts on the server, to achieve a superior trade-off between the global generalization and the local personalization. To demonstrate the effectiveness of our PGA-FL framework, we theoretically prove its convergence performance and perform extensive experiments. Experimental results demonstrate that the proposed PGA-FL framework can achieve superior performance in terms of convergence performance, global generalization, and local personalization with respect to other state-of-the-art baseline algorithms.

Low Earth Orbit satellites play a vital role in Earth observation and remote sensing missions, with their satellite-to-ground downlinks responsible for transmitting substantial volumes of imagery and data. However, the inherent dynamism and instability of satellite-to-ground links present challenges to efficient data transmission. While existing research often concentrates on predicting physical layer parameters, our empirical measurements reveal that solely relying on physical layer predictions is insufficient to reflect application layer throughput accurately. To address this, we propose Satformer, a spatio-temporal prediction model, and based on it, the Spatio-Temporal cross-layer Adaptive Rate control (STAR) mechanism to achieve efficient satellite-to-ground downlink transmission. To bridge the gap between physical and application layers, Satformer employs tailored Sat-Embedding and a spatio-temporal attention mechanism that explicitly correlates satellite relative position data and weather data with physical layer data to predict application layer throughput. Experimental results show that Satformer improves application layer throughput prediction accuracy by up to 33.45% compared to baseline models. Building on this, the STAR mechanism increases satellite-to-ground data transmission throughput by an average of 12.12%, while reducing median end-to-end latency by up to 43.3%.

The rapidly growing edge computing platforms, coupled with the still imperfect edge infrastructure, present an excellent opportunity for the emergence of new edge hardware. However, it remains unclear whether alternative architectures built from energy-efficient mobile System-on-Chips (SoCs) can meet the stringent performance, cost, and energy demands of modern edge workloads. In this paper, we propose a new type of edge server composed of 60 Qualcomm Snapdragon 865 mobile SoCs in a 2 U rack, referred to as SoC Cluster. We demonstrate its successful deployment on existing edge cloud platforms and its ability to natively serve mobile cloud gaming services. Despite the emergence of new hardware on edge platforms and its successful operation in serving mobile cloud gaming, our trace analysis revealed low hardware utilization and significant dynamic fluctuations in usage. To assess its broader applicability, we conducted the first measurement study of SoC Cluster to reveal its ability to run two popular and modern edge applications: deep learning inference and video transcoding. We developed a cross-platform benchmark suite to evaluate throughput, latency, power consumption, and application-specific metrics like video quality. We then directly compare SoC Cluster with a traditional edge server equipped with Intel CPUs and NVIDIA GPUs in terms of energy efficiency, space efficiency, and monetary cost. Results show that SoC Cluster exhibits up to 6.5 & times; higher energy efficiency and 7.7 & times; higher space efficiency. We also disclose its limitations in serving computation-intensive workloads such as large deep learning models. The outcomes provide insightful implications and offer practical direction for refining SoC Cluster toward broader deployment in edge scenarios.

The rapid growth of mobile applications (particularly deployed natural language processing applications), generate a proliferation of user-generated data, encapsulating a wealth of valuable knowledge. To facilitate the analysis and extraction of insights from on-device data while adhering to stringent data privacy regulations, natural language processing (NLP) models deployed on devices are frequently finetuned within a federated learning (FL) setting. However, this process typically requires hundreds of thousands of labeled training samples from mobile users. This presents a significant challenge, as users often lack the motivation or expertise to annotate their data. The scarcity of labeled data, referred to as the few-shot scenario, poses a substantial obstacle for mobile NLP applications. This work is the first to explore federated NLP in the few-shot scenario (FedFSL). Extensive measurements are conducted to highlight that previous federated NLP suffer from significant accuracy degradation when labeled data is scarce and imbalanced distributed. By leveraging advancements in pseudolabeling and prompt learning, we establish a training pipeline that can achieve competitive accuracy on the evaluated fewshot NLP benchmarks with as few as 0.05% of training samples (fewer than 100) labeled, while the rest remain unlabeled. To make this pipeline into practice, we introduce a system, FeS1, which addresses high execution costs through three key innovations: (1) Curriculum pacing, which injects pseudo labels into the training process in alignment with the learning progress; (2) Representational diversity, a strategy that selects the most learnable data for pseudo-label generation; (3) Coplanning of the model's training depth and layer capacity. Together, these designs reduce training latency and client energy consumption by up to 46.× and 41.2×, respectively. For network traffic, FeS achieves an average reduction of 1841.7× over FedFSL and 2.17× over the stronger FedFSLBIAS baseline across four tasks. By integrating algorithmic and system-level innovations, FeS demonstrates that FL can be effectively applied in challenging scenarios where the majority of training samples are unlabeled.

Retrieval-Augmented Generation (RAG) services enhances LLMs by dynamically retrieving relevant information from external knowledge sources, addressing their limitations in handling up-to-date or domain-specific knowledge. Under the constraints of the cloud-native LLM service paradigm, many existing RAG systems are typically deployed in cloud environments and rely on retrieving raw information from local datasets of users or third-party knowledge providers to enhance generation. However, these cloud-based RAG services often neglect critical privacy and intellectual property concerns of the knowledge base, potentially leaking sensitive information during deployment or inference. To address these risks, we propose a privacy preserving framework (PriRAG) via proxy knowledge for cloud based RAG. The core idea is to compress the raw data of the retrieved information into proxy knowledge representations with carefully-designed privacy preservation mechanisms. We first propose a knowledge construction mechanism that aligns proxy representations with the semantic space of LLMs to preserve utility. Then we develop a query-centric knowledge representation method with a tailored training strategy to selectively exclude the irrelevant content, further reducing the exposure of privacy sensitive content. We further formulate a differential privacy enhanced collaborative inference strategy to secure the whole framework with certifiable guarantees. Extensive experiments on 7 public benchmarks and 4 LLMs show that PriRAG achieves a favorable balance between utility and privacy, significantly reducing privacy risks while maintaining high-quality generation.

Large language models (LLMs) are deployed on mobile devices to power killer applications such as intelligent assistants. LLMs pre-trained on general corpora often hallucinate when handling personalized or unseen queries, leading to incorrect or outdated responses. Knowledge editing addresses this by identifying and adjusting a small crucial portion of model weights, without compromising the general knowledge. However, prior knowledge editing methods are impractical to run on local devices due to the resource-heavy backpropagation (BP) needed for updates. We present MobiEdit, the first mobile knowledge editing framework that enables efficient LLM personalization on commercial off-the-shelf (COTS) mobile devices. MobiEdit replaces full-precision BP with quantized forward-only gradient estimation, thus compatible with the energy-efficient mobile neural processing units (NPUs). To further improve gradient estimation efficiency, we introduce two optimizations: an early stopping mechanism that adaptively terminates editing upon success and prefix activation reusing that reduce redundant computation across steps. Our approach enables real-time editing of 3B-parameter models (Qwen2.5-3B-Instruct and Llama3.2-3B-Instruct) on COTS mobile devices with 7.1$\times$ less memory, 15.8 $\times$ less energy and 3.4$\times$ less latency compared to previous knowledge editing methods.

Satellite computing, as an emerging edge computing paradigm, extends computing and networking services into space. Due to the internal design constraints of low-Earth orbit (LEO) satellites and the challenges posed by the external environment, satellite computing faces inherent limitations, including severely constrained resources, non-rechargeable batteries, poor heat dissipation, and highly dynamic operating conditions, leading to unreliable and unsustainable quality of service. To address the above challenges and fully realize the potential of satellite computing, this paper investigates temperature- and energy-aware dynamic task scheduling and computing resource allocation, aiming to optimize service latency, reduce onboard energy consumption, and enhance operational profit. Solving this problem requires coordinating task scheduling and resource allocation, balancing communication and computation latency, and addressing the challenge of a vast search space. To solve the above challenges, we first formulate this problem as a repeated Stackelberg game by developing temperature and energy models. Through theoretical analysis, we show that this game leads to a convex optimization framework that exhibits exponential complexity. To accelerate the search for the Stackelberg equilibrium solution, we propose a dynamic task scheduling algorithm based on the interior point method, which reduces the computational complexity to polynomial order. Trace-driven simulations demonstrate that the proposed algorithm reduces task scheduling latency by 28.4% and improves utility by 13% on average.

With the rapid growth of service computing, sparsely activated large language models (LLMs), particularly those based on Mixture-of-Experts (MoE), have emerged as a promising foundation for large-scale service deployment. Their key advantage is lower service inference cost with high model capacity. However, before deployment, these models generally require fine-tuning to improve downstream service quality. In many applications, the fine-tuning data are distributed across edge clients and cannot be centrally collected due to privacy, regulatory, and ownership constraints. This makes federated learning a natural paradigm. Nevertheless, federated fine-tuning of MoE-based LLMs is challenging. Their large size imposes prohibitive computation and memory costs on resource-limited clients. Meanwhile, client heterogeneity and non-IID data further hinder efficient training. Existing optimization strategies, such as parameter-efficient fine-tuning, knowledge distillation, and expert offloading, either sacrifice performance or lack practicality in federated settings. To fill this gap, we propose FedMoE, a federated fine-tuning framework for MoE-based LLMs. FedMoE constructs a compact client-specific sub-MoE tailored to each client's data characteristics and resource constraints. FedMoE integrates three components: expert preference profiling for efficient estimation of clientspecific expert activation patterns, resource-adaptive sub-MoE orchestration for dynamic expert selection and modular aggregation, and cross-client collaborative expert recommendation for discovering useful experts beyond the local sub-MoE. Extensive experiments across multiple federated settings show that FedMoE consistently achieves superior training effectiveness and efficiency over strong baselines under heterogeneous conditions, while also demonstrating strong robustness and stable performance across diverse settings.

Cloud-assisted resource-constrained and widely used outdoor IoT devices deploying highly robust convolutional neural network (CNN) models to provide high-quality visual services have attracted widespread attention from industry and academia. Most existing methods suffer from two limitations: (i) large amount of parameter transmission, and (ii) low robustness in handling data affected by the environment. To this end, this paper proposes a transmission-friendly and robust CNN design method called NonLinear transformation generation Filter with NonLearnable hyperparameters, namely NL2Filter. In NL2Filter, some filters are first designated as seed filters, whose parameters are learnable, that is, updated as the model is trained. Other filters in this layer are generated based on the seed filters and nonlinear transformation function (NLF). The hyperparameters of the NLF in NL2Filter are randomly initialized and remain unchanged, so they can be saved and reproduced using random seed. After the cloud server trains the NL2Filter-CNN model, it only needs to send a small number of learnable parameters and random seed to reproduce the CNN model trained by the cloud server. Compared with sending the complete CNN model, sending only a small number of learnable parameters and random seed can significantly reduce the number of model parameters sent by the cloud server. On the other hand, NL2Filter groups seed filters and then uses NLFs with different hyperparameters to generate diverse filters on demand for each group, thereby improving the model's ability to capture feature diversity without increasing the complexity of the model, thereby improving the model's robustness in processing data affected by the environment. Experimental results on CIFAR-10, CIFAR-10-C, and Icons-50 datasets demonstrate that the proposed NL2Filter outperforms other state-of-the-art methods. Specifically, using the ResNet-101 architecture, on the CIFAR-10-C, NL2Filter's accuracy is about 3.9% higher than that of MonoCNN; On the Icons-50, NL2Filter's accuracy is about 2.7% higher than that of MonoCNN and about 2.3% higher than that of the standard ResNet-101.

Containerization offers great flexibility and agility for resource provisioning in edge clouds. However, this benefit is not freely available, as substantial network traffic incurred by container image pulling heavily burdens the back-haul networks. Our in-depth measurements on 516.3 GB images from Docker Hub reveal that many similar images share identical layers, with only 11.21% having no shared layers. Building on this, we investigate the resource provisioning problem by leveraging image similarity to avoid repeated layer transmissions, aiming to reduce network traffic and improve overall performance. We formulate resource provisioning as a mixed-integer non-linear programming problem, which is challenging due to the coupling of four issues and their conflicting effects on overall performance, including offloading decisions, container instance deployment, image pulling, and resource allocation. To tackle these complexities, we propose a novel Similarity-Aware Resource Provisioning approach, which decomposes the problem into independent sub-problems using counterfactual multi-agent deep reinforcement learning and then solves sub-problems individually with convex optimization and fractional programming techniques. We conduct extensive evaluations with images from Docker Hub. The results show that our approach enables up to 32.6% traffic reduction and 19.9% utility improvement, outperforming the state-of-the-art solutions.

Low Earth orbit satellites are emerging as critical computing platforms for real-time environmental monitoring applications, yet on-orbit satellites face dual risks from space radiation-induced hardware failures and load-driven degradation. Task migration is key to maintaining service resilience. However, existing migration methods are limited in jointly optimizing migration latency and load balancing given their inherent coupling. They also lack robustness to sudden satellite risk changes and link fluctuations during migration. Therefore, we propose SatRTM, a satellite risk-aware task migration system that incorporates a dual-component fault risk model to quantify migration risk from radiation and load-driven degradation, and formulates the migration as a dual-objective optimization minimizing migration latency while maximizing mean time between failures (MTBF). Accordingly, we model inter-satellite migration interactions as a coalition game and propose the coalition-based optimal migration scheduling algorithm to compute the equilibrium, provably yielding the optimal migration strategy. To sustain migration strategy validity, we propose a dynamic robust coalition reconfiguration algorithm that achieves adaptation by two novel components, migration progress tracking and risk event-triggered coalition adjustment. Extensive experiments show that SatRTM reduces service recovery latency by 35.6%, improves load balance by 14.3%, and extends satellite MTBF by 17.3% over baselines. Under uncertain migration conditions, SatRTM further achieves up to 82% faster fault recovery and sustains an 84.2% task completion rate under bandwidth degradation.

Cross-datacenter AI training often suffers from WAN packet loss and high retransmission latency. Existing FEC schemes for RDMA transmission lack application awareness, struggling to balance recovery reliability and bandwidth overhead. We propose Training-aware Loss Recovery(TLR), a domain-specific loss recovery scheme that exploits the asymmetric loss tolerance bound of model gradients. TLR prioritizes critical gradients with high redundancy and employs quantization for FEC inputs to minimize bandwidth consumption. Evaluations show that TLR significantly outperforms state-of-the-art schemes in both average and tail Flow Completion Times (FCT) performance.

The growth of large models demands multinode cooperation during training and inference processes. The computing node failures can interrupt these processes, subsequently causing information loss and prolonging the execution time. To reduce the prohibitively large overhead incurred by the computing nodes failures, the accurate prediction of computing node failure is vital, which can help to avert potential large overhead, service interruptions, and negative customer experiences. Existing solutions of computing nodes failure prediction mainly focus on utilizing state-of-the-art time-series models to enhance the performance of computing node failure prediction. However, on the one hand, they could not capture the causal relationship between device over-utilization and node failures; On the other hand, they fail to extract the complex spatial-temporal cascading correlations among computing node failure events. These limits can degrade the performance of computing node failure prediction. To address these above problems, this article makes an effort to focus on designing a continuous-time dynamic graphs-based computing node failures prediction (CTDG-NFP) scheme, to accurately predict in dynamic cluster environments. Specifically, the CTDG-NFP scheme first designs a novel multiple-dimensional feature-biased neighbor sampling method, which jointly considers CPU utilization-biased, memory utilization-biased, temporal-biased and spatial-biased, to sample relevant context. Then, the CTDG-NFP scheme extracts diverse computing node failure motifs by multiple-dimensional feature-biased-based long-short-path walk method and set-based anonymization method. Finally, the CTDG-NFP scheme adopts time encoder to encode these motifs, and thereby extracting the complex spatial-temporal correlations among computing node failure events. On this basis, contrastive learning is adopted to train the computing node failure prediction model. Extensive evaluations with various real-world failure traces demonstrate the CTDG-NFP scheme can achieve superior performance in terms of six widely used performance metrics compared with the SOTA node failure prediction methods.

Satellite computing enhances Earth observation, disaster monitoring, and weather forecasting through in-orbit data processing, yet the harsh space environment exposes satellite nodes to stochastic failures, making service migration critical for proactive state transfer and service continuity. However, satellites within the same orbital plane share identical radiation environments, exposing migration destinations to correlated fault risks. Meanwhile, strictly constrained onboard resources make overloaded migration destinations prone to becoming new failure sources, triggering cascading failures. Existing approaches fail to jointly optimize migration latency and load balance under dynamic fault awareness, leaving cascading failure risks unresolved. To fill this gap, this paper proposes SatBaco, an adaptive satellite service migration framework that formulates migration target selection as a coalitional game and proves Nash equilibrium is equivalent to the joint latency-load optimal solution. Migration node coalition formation algorithm is then proposed to resolve conflicting interests between at-risk satellite and candidate nodes, deriving the optimal migration targets. To further adapt to continuously evolving node health states, a dynamic coalition adjustment algorithm guided by risk-aware neighborhood selection is proposed, which dynamically updates migration decisions according to emerging fault warnings and changing node reliability. SatBaco jointly optimizes migration latency, load balance, and dynamic fault risk, preventing cascading failures across the network. Extensive evaluations driven by realistic constellation information have demonstrated excellent performance of SatBaco, achieving 39.05% migration latency reduction and 26.34% satellite system average load pressure decrease compared to existing methods.

To accommodate the diverse requirements of coexisting AI applications and background workloads, datacenter networks often employ hybrid load balancing strategies: Equal-Cost Multi-Path (ECMP) for background flows and Random Packet Spraying (RPS) for AI traffic. However, in this paper, we show that ECMP-induced hotspots create severe path-delay asymmetry, leading to excessive packet reordering for RPS flows. This reordering is misinterpreted by the RoCEv2 transport layer as congestion signals, causing unnecessary rate reduction and significant throughput degradation. To address these problems, we propose ynamic Load Balancing Framework (DLBF), an end-host load-balancing mechanism. DLBF dynamically migrates traffic away from hotspots based on per-path congestion signals. Further, by transmitting duplicate packets based on the delay gap between old and new paths, DLBF eliminates reordering during path switching without violating Remote Direct Memory Access (RDMA) semantics. Last, DLBF utilizes lightweight notifications to probe path states without aggravating congestion. Large-scale NS-3 simulations demonstrate that DLBF reduces the average Flow Completion Time (FCT) by 1.1×–1.7× and lowers P99 tail latency by 1.2×–3.4× compared to existing schemes.

Physical AI policies require inference throughout their lifecycle, including model evaluation, cloud reinforcement learning rollout, edge GPU serving, and onboard deployment. Although these settings share the same checkpoint and action semantics, they often rely on separate inference programs. To unify them, we build PhyAI, a Physical AI inference engine with a single runtime that keeps architecture-specific conditioning, solver, cache, and output logic in model adapters while sharing graph execution, kernels, memory management, and parallel services. The same codebase runs vision-language-action (VLA) models and world-action models (WAMs) on single or multiple GPUs across onboard, edge, and cloud deployments. We used the adapter interface to add MiniCPM-Robot on the day of its release. PhyAI achieves 1.40x-4.65x speedups over the official implementations of pi0, pi0.5, GR00T N1.7, and MiniCPM-Robot. On Cosmos3-Nano-Policy-DROID it reduces latency from 2.46 to 1.18 s on eight H20 GPUs (CFG=2, TP=4), a 2.08x speedup. Specialized runtimes remain faster in several configurations, so our goal is one runtime with competitive latency rather than the fastest result in every case. Detailed profiles reveal why different models need different execution policies: on a Hopper-series GPU at batch size one, the pi0.5 action expert accounts for 8.8

Using low-rank adaptation to fine-tune pre-trained neural network models has attracted widespread attention due to its advantages of low resource requirements, high precision, and no additional inference delay. However, most existing methods are designed for large language models based on the Transformer structure and lack adaptation to convolutional neural networks (CNN) widely used on Internet of Things (IoT) devices. In addition, IoT devices are usually deployed outdoors and collect large amounts of data that are affected by environmental conditions. Providing a highly robust CNN model is a prerequisite for providing high-quality services. To this end, this paper proposes a highly robust seed-filter adaptation method (SeFA) for pre-trained CNNs. SeFA introduces an adaptation branch with the same structure as the backbone network. In the adaptation branch, some filters are first designated seed filters and grouped. Then, additional filters are generated from the grouped seed filters and nonlinear transformation functions (NLFs) with different hyperparameters. The seed filters' parameters are updated during model training, and the NLF hyperparameters are randomly initialized and frozen. Both grouping seed filters and configuring NLFs with non-learnable hyperparameters can improve the model's robustness. This is because grouping seed filters can generate diverse filters on demand without increasing the model's complexity, and the NLFs' rules can regularize the model. It is worth mentioning that the number of fine-tuning parameters of the pre-trained model that can adapt to downstream tasks can be flexibly controlled by specifying the number of seed filters. The key idea of this paper is to propose SeFA with flexible, controllable, learnable parameters, high robustness, and adaptability to pretrained CNN models, thereby facilitating fine-tuning on resource-constrained IoT devices and providing highly robust visual services. Experimental results on the CIFAR-10, CIFAR-10-C, CIFAR-100, CIFAR-100-C, and Icons-50 datasets demonstrate that the proposed SeFA outperforms other state-of-the-art methods. Specifically, on the ResNet-152 model and the CIFAR-10-C dataset, the accuracy of our SeFA is about 7% higher than that of the full fine-tuning method. In addition, we verify the efficiency of SeFA-based models for fine-tuning and inference on the device.

Federated map matching aims to align GPS trajectories collected by multiple decentralized clients to road networks in a privacy-preserving manner, and remains an unexplored yet critical problem for collaboration across clients. However, this task is highly challenging due to severe client heterogeneity, which hinder efficient and robust model convergence. This paper presents a novel federated learning framework that achieves high accuracy, communication efficiency, and robustness in heterogeneous environments. At the model-level, we design Gel-HMM, a graph-enhanced, learnable Hidden Markov Model that decouples client-specific information through specific graph structures and supports interpretable matching. At the framework-level, we propose a curriculum-aware client selection strategy that captures global learning dynamics and adaptively guides client participation. Furthermore, we develop a dual-speed parallel optimization scheme to decouple the co-training of the client matching model and the server-side selector. An adaptive knowledge distillation mechanism is also introduced to mitigate data quality variance across clients while preserving privacy. Extensive experiments on two real-world datasets demonstrate that our method significantly outperforms state-of-the-art baselines, with an average improvement of 20% on accuracy and a 60% reduction in communication cost.