On behalf of the technical program committee, we are delighted to invite you to participate in the 2025 IEEE 16th International Conference on Cloud Computing Technology and Science (IEEE CloudCom), to be held in Shenzhen, China on November 14-16, 2025.
Building automation is a representative AIoT-driven cyber-physical scenario, where intelligent systems interact with physical devices to manage lighting, climate, and appliances in real time. Traditional machine learning struggles with ambiguous, multilingual, and colloquial user inputs, limiting effectiveness in dynamic building environments. Recent advances in large language models (LLMs) enable more natural command interpretation, but high resource demands challenge their sustainable deployment on edge nodes. This paper proposes a serverless architecture based on event-driven microservices and container orchestration that dynamically manages the deployment, execution, and scaling of compact, fine-tuned LLMs across distributed edge nodes for building automation. We fine-tune compact LLMs with ambiguous and colloquial command examples to enhance robustness and enable context-aware deployment at the edge. Platform elasticity, enabled by Knative, allows rapid model adaptation without persistent resource allocation. We evaluate the system on a multilingual building automation dataset (Chinese, English, French) with ambiguous and colloquial commands, using an automated framework to assess interpretation and execution. Results show that the fine-tuned model outperforms the baseline Qwen-2.5-14B on five of six metrics and performs comparably on output format compliance. It also generalizes well across languages, although fuzzy instructions remain challenging.
Mobile Personal Livestreaming (MPL), one of the most popular mobile applications recently, has been experiencing explosive growth thanks to ubiquitous smartphones and high-speed wireless connections nowadays, e.g., 4G-LTE/5G and WiFi 802.11ac/ad. Unlike traditional livestreaming services that are either limited to specific content providers or fixed content, MPL enables end-to-end live-stream communication of anything, at anytime and anywhere, with anybody in mobile networks. This paper presents our design and experience of Inke, a large-scale MPL system with 15 million daily active users since 2015, which provides a range of functionalities, including mobile livestreaming, real time communication, augmented reality (AR), and real time face recognition. We characterize its system design from three key aspects: push-streaming protocol, “Join Channel“ livestreaming, and AR features. For the push-streaming protocol, we introduce the bandwidth-aware design that incorporates buffer length control and frame-drop strategy to ensure upload consistency and reduce delay. We also design a mixed-stream upload model that is able to balance the delay and bandwidth consumption during “Join Channel” livestreaming. In addition, we show how the virtual gift value distribution can be used to im-prove the design of animation playback queues. We demonstrate the effectiveness of our system design that works across time, geo-location, and heterogeneous wireless networks for large-scale (more than 130 million registered users) applications. It achieves 126.5736ms average upload delay, 1.093 average rebuffering times during each broadcast at peak time.
The success of distributed machine learning and federated learning in cloud environments depends critically on efficient communication across heterogeneous and bandwidth-constrained networks. While the exchange of raw data is often avoided due to privacy and regulatory concerns, many distributed workflows still require large-scale data movement or model updates that demand predictable network performance. This paper investigates the trade-off between earliest completion time (ECT) and communication cost in data transfers that leverage bandwidth reservation on fixed paths with variable bandwidth in dedicated high-performance networks. We fo-cus on two representative scheduling objectives relevant to distributed learning and data-intensive cloud workflows: (i) minimizing communication cost while meeting a transfer dead-line, and (ii) achieving the ECT while respecting a maximum cost budget. We prove that both scheduling problems are NP-complete and propose heuristic algorithms to address them effectively. Extensive simulations demonstrate that our methods can substantially reduce both latency and cost, providing a prac-tical foundation for communication-efficient and resource-aware distributed machine learning in cloud computing environments.
Satellite communication plays a crucial role in future networks, but traditional systems face significant challenges in interference management and energy efficiency. To address these issues, this paper proposes a green satellite communication framework that combines Lyapunov optimization with a one-dimensional golden-section search method. The framework builds a two-timescale frame-slot model that jointly captures fast-varying channel fading and slow-varying renewable energy dynamics. By introducing a drift-plus-penalty optimization method, the system ensures queue stability and minimizes long-term grid energy expenditure, while employing the golden-section search to optimize beamforming parameters with reduced computational complexity. Simulation results show that the proposed framework effectively balances energy efficiency and communication performance, demonstrating good scalability for large-scale satellite networks.
With the rise of Web3, Non-Fungible Tokens (NFTs) have become a new class of digital assets, driving demand for large-scale NFT recommendation systems. Each NFT can be associated to a rich set of semantic, stylistic, and thematic labels, forming a highly complex label space. Similar to e-commerce platforms where detailed product labels enable personalized recommendations, such semantic dependencies between labels can potentially enhance NFT recommendation performance. Thus, NFT recommendation can be naturally formulated as an extreme multi-label (XML) classification problem. Many existing probabilistic label tree (PLT)-based approaches address XML problem by recursively partitioning the label space, which greatly alleviates the demands on expensive computer resources. Yet, the highly skewed distribution of labels in datasets in XML makes tail labels more challenging to predict than head labels. In this paper, Our preliminary analysis reveals that inherent label dependencies can be leveraged to improve tail label recommendations for NFTs. We propose ChainTail, a dependency-aware framework that enhances PLT-based NFT label partitioning and prediction re-scoring. It includes: (1) a Dependency-aware partition module that partitions highly dependent NFT labels into subsets. (2) a Dependency-aware ReScore module that re-ranks prediction scores of labels to eliminate the label-priors. Our experimental results show that ChainTail boosts tail label recommendation on widely used item recommendation datasets.
Large Language Models (LLMs) have achieved remarkable success but face significant deployment challenges in cloud and edge environments due to their massive computational and storage requirements. Model quantization serves as a key solution to enhance the scalability and efficiency of LLMs within distributed cloud platforms. Existing Post-Training Quantization (PTQ) methods often exhibit suboptimal performance in low-bit settings. To further improve their precision, Quantization-Aware Training (QAT) combined with Low-Rank Adaptation (LoRA) has been explored for error correction. However, a critical issue is that the quantized base model and full-precision LoRA parameters suffer from precision mismatch, introducing additional errors during weight merging. To address these challenges, we propose a Quantized Low-rank Error Reconstructor (QER) for LLM low-bitwidth quantization. QER first enables lossless merging in low-bitwidth format by aligning the bitwidth of its low-rank parameters with the quantized base parameters, eliminating dequantization and requantization steps. Through this process, QER reconstructs original errors into two components: the quantization errors of QER parameters (i.e., quantized low-rank parameters) and potential overflow errors during low-bitwidth merging. These two errors are directly related to QER parameters, making them easier to optimize via gradient-based updates within an error-aware training framework. Requiring only 128 samples and 1 training epoch, QER demonstrates superior performance on LLaMA-1/2 families. In 4-bit quantization, compared to QLLM with error correction, QER reduces average perplexity by 13.8% (from 10.97 to 9.45) and improves average accuracy by 3.01 percentage points (from 51.84% to 54.85%) on LLaMA-1-7B. QER bridges the gap between quantization and low-rank adaptation, enabling efficient and accurate low-precision LLM deployment.
Cloud-native technologies have become a corner-stone of modern enterprise computing, yet they continue to face two major challenges compared to traditional software development: complex configuration and operations, and excessive resource consumption. These limitations often hinder adoption by small and medium-sized enterprises as well as individual developers. In this paper, we present CloudAtion, a functional programming framework for cloud-native container resource management. CloudAtion introduces functional abstractions into container image construction and runtime orchestration, enabling efficient resource reclamation and reduced operational overhead. In addition, the framework provides a visual programming interface to simplify configuration and reduce development complexity. Experimental evaluations show that CloudAtion achieves faster and more efficient image and container recycling compared with conventional Kubernetes and Docker mechanisms, while also reducing redundant complexity in Dockerfile specifications. Overall, the proposed framework improves runtime efficiency and offers a systematic and controllable solution for resource management in cloud-native environments, providing both theoretical insights and practical guidance for future container orchestration and optimization.
Severe class imbalance in network traffic data poses a significant challenge to intrusion detection systems (IDS), particularly in identifying rare attack categories. Although extensive research has proposed various enhancements for minority attack detection, the effectiveness of different preprocessing strategies varies greatly between attack types, and some approaches even degrade detection performance for certain classes. In this study, we conducted a systematic evaluation using the CICIDS2017 dataset, designing six distinct data processing schemes and integrating six supervised learning models. Random Forest (RF), Decision Tree (DT), Bayesian Network (BN), Naive Bayes (NB), Support Vector Machine (SVM) and Deep Neural Network (DNN). Our comprehensive experiments assess the detection capabilities for each attack type and analyze the variability in the generalization performance of different strategies across diverse attack scenarios. The results reveal the applicability and limitations of common preprocessing strategies in highly imbalanced cloud environments, providing practical insights for IDS model optimization and feature selection in real-world cloud computing contexts.
In today's highly competitive market environment, manufacturing enterprises face critical challenges such as difficulties in quantifying customer credit risk, unclear customer value differentiation, and unpredictable customer churn. Aiming to address these challenges, this paper proposes IntelligentCRM, a data-driven Customer Relationship Management (CRM) system designed for the manufacturing industry. The system integrates three core analytical models: firstly, a comprehensive credit evaluation model that accurately assesses customer credit through a multi-dimensional indicator system; secondly, a customer segmentation model that effectively categorizes customer value based on transaction behavior and other data; and finally, a customer churn prediction model that proactively identifies at-risk customers using machine learning techniques. This paper elaborates on the system's overall architecture, the design and implementation of its core modules, and demonstrates its feasibility through model construction and business process analysis. IntelligentCRM provides a complete solution for enterprises to transition from passive management to proactive and intelligent customer relationship management, ultimately aiming to enhance risk control capabilities, optimize resource allocation, and improve customer loyalty.
We introduce a self-supervised, Multi-Task Framework that simultaneously recognises human emotion and synthesises emotion-consistent motion within a single skeleton-based networkan ability that is particularly valuable for resource-constrained Internet-of-Things (IoT) devices that must both perceive and respond in situ. A masked-motion auto-encoder with factorised space-time attention first reconstructs hidden bodypart tokens to mine fine-grained spatio-temporal cues without labels; the same encoder then feeds a lightweight classification head that predicts eight Ekman-style emotions. Evaluated on the Emilya benchmark (8206 clips, 8 daily actions × 8 emotions), our model attains 88.7% accuracy and 0.886 macro-F1, matching state-of-the-art recognition-only GCNs while also generating realistic, emotion-specific motions. Visualisations show expansive, dynamic poses for joy and contracted, slow movements for sadness; ablations reveal that the joint reconstruction loss and factorised attention contribute roughly 2 pp and markedly boost minority classes such as pride and shame. Because the classifier is ultra-compact and the encoder can be quantised or partitioned, the framework prepares itself for future deployment on low-power IoT hardware, enabling edge devices, robots, and wearables to sense and express affect in real time.
Container orchestration platforms are essential for deploying microservices, yet their impact on security-intensive workloads remains largely misunderstood. This work presents a systematic performance comparison of SPIFFE-based identity artifacts across Docker Compose and Kubernetes, evaluating four identity modes through fine-grained instrumentation of token minting and validation operations. Our results reveal orchestration overhead depends critically on connection management: operations employing connection pooling exhibit minimal overhead (0.8-4.9 %), while repeated connection establishment incurs substantial penalties (up to 26.6%). Kubernetes demonstrates 2-10x higher performance variability than Docker Compose, indicating platform selection affects predictability beyond mean latency. These findings, obtained using minimal Container Network Interface (CNI) configuration, represent a conservative baseline for production deployments and provide empirical guidance for platform selection and security architecture design.
Vision language models (VLMs) have opened new avenues for satellite remote sensing image analysis and have shown promise across multiple tasks. However, in the absence of a remote sensing-oriented general VLM, existing approaches rely on retraining generic VLMs with remote sensing datasets to adapt to downstream tasks. This practice is inherently affected by two factors: 1) generic VLMs are pretrained on massive web data containing noise, biases, and misinformation; and 2) many remote sensing image-text datasets use VLM-generated annotations, which can introduce hallucinations and factual errors. Such statistical biases exacerbate the alignment gap of remote sensing VLMs, leading to generation bias and degraded task performance. To address this issue, we propose Geo-Contrastive Decoding (Geo-CD) to enhance remote sensing VLMs. Geo-CD reduces over-reliance on statistical biases by contrasting the output distributions produced from distorted versus original visual inputs. This strategy ensures that the generations of VLMs remain well-grounded in the visual input, thereby improving both reliability and accuracy. Extensive experiments demonstrate that Geo-CD can be applied to diverse remote sensing tasks without additional training or external tools. On the selected base VLM, Geo-CD achieves consistent gains across most remote sensing benchmarks and reaches state-of-the-art performance.
This study develops a BO-driven dual-factor quantitative trading strategy integrating optimized XGBoost/LSTM models to address market nonlinearity and volatility. The BO-optimized XGBoost model shows superior accuracy and, when combined with a 9-day moving average, yields a 20.79% annualized return and a 1.14 Sharpe ratio. The strategy offers superior signal filtering, risk-adjusted returns, and robustness, laying a foundation for adaptive trading systems.
The growth of distributed energy resources and local energy markets heightens the need for price formation that is transparent, privacy preserving, and compatible with network constraints. Blockchain provides a trust-minimized substrate for auditable clearing and settlement through consensus, tamperevident ledgers, and smart contracts. This survey organizes blockchain-enabled pricing into three families, namely auction-based, game-theoretic, and optimization-based, and links them to enabling techniques such as metering oracles, secure multiparty computation, zero-knowledge proofs, and verifiable optimality certificates. Applications span wholesale electricity, carbon and green certificates, distributed energy trading, ancillary services, and electric vehicles. Evidence indicates gains in auditability, privacy, network awareness, and automated settlement, alongside challenges in scalability, data protection, grid integration, and regulation. The survey distills design patterns and research directions toward verifiable, interoperable, and governable pricing modules that complement system-operator markets.
The malicious misuse of deep learning-based video inpainting techniques poses significant security risks, highlighting the critical importance of accurately detecting inpainted regions in digital video content. However, existing methods suffer from limited accuracy in complex motion scenarios, and their cross-modal feature fusion efficiency is low due to simplistic integration strategies. To address these challenges, we propose FAD-Net, an end-to-end dual-branch framework that performs video inpainting localization by jointly modeling optical flow and RGB modalities to complement each other's capabilities. The optical flow branch employs a flow consistency error constraint and flow anomaly awareness module to mitigate the impact of inaccurate optical flow estimation, while the parallel RGB branch utilizes an encoder for spatial texture feature extraction and cross-frame attention for long-term temporal modeling. A bidirectional dynamic fusion module then adaptively integrates complementary features using motion-aware weights. Experimental results demonstrate that FAD-Net outperforms existing methods in accurately localizing the inpainted regions, particularly exhibiting excellent performance in complex dynamic scenarios and with unknown inpainting types, enabling reliable forensic analysis of tampered videos.
The rapid growth of multimodal data in 6G-enabled applications demands a paradigm shift from traditional bit-rate-oriented communication to semantic transmission. While diffusion models offer a promising solution for high-fidelity generation, their application is hindered by a fundamental disconnect from the physical channel and high computational overhead. Existing methods typically treat the denoising process as independent of the channel conditions, using abstract integer timesteps that lack physical meaning, and require a full, lengthy sampling chain for reconstruction. To address these challenges, we propose signal-to-noise ratio-guided latent diffusion model (SNR-LDM), which re-parameterizes the diffusion process to map denoising time steps directly to physical SNR. This enables adaptive inference from noise levels precisely matched to the estimated channel state, significantly reducing redundant computations while further guiding the reconstruction process through multimodal prompts to ensure semantic consistency. Experimental results demonstrate that, across a wide range of channel conditions, our proposed SNR-LDM achieves a PSNR gain of approximately 4 dB over ADJSCC while simultaneously reducing the LPIPS from 0.75 to 0.25.
Distributed optimization is pivotal for large-scale signal processing and machine learning, yet communication overhead remains a major bottleneck. Low-rank gradient com-pression, in which the transmitted gradients are approximated by low-rank matrices to reduce communication, offers a promising remedy. Existing methods typically adopt either randomized or greedy compression strategies: randomized approaches project gradients onto randomly chosen subspaces, introducing high variance and degrading empirical performance; greedy methods select the most informative subspaces, achieving strong empirical results but lacking convergence guarantees. To address this gap, we propose GreedyLore-the first Greedy Low-Rank gradient compression algorithm for distributedlearning with rigorous convergence guarantees. GreedyLore incorporates error feedback to correct the bias introduced by greedy compression and introduces a semi-lazy subspace update that ensures the com-pression operator remains contractive throughout all iterations. With these techniques, we prove that GreedyLore achieves a convergence rate of $\mathcal{O}(\sigma/\sqrt{NT}^{-}+1/T)$ under standard optimizers such as Adam-marking the first linear speedup convergence rate for low-rank gradient compression. Extensive experiments are conducted to validate our theoretical findings.
Digital Twins (DTs) are increasingly deployed in collaborative ecosystems, enabling adaptive monitoring, simulation, and decision-making across interconnected systems. Similar to Internet of Things (IoT) systems, DTs rely on real-time data collection, analysis, and feedback loops to represent and manage physical entities. Both technologies generate large volumes of data and leverage AI/ML techniques to optimise operations, making them complementary in building intelligent, adaptive ecosystems. However, the heterogeneous and dynamic nature of these ecosystems introduces challenges in assessing trustworthiness and ensuring reliable collaboration. This paper presents a Trust Analyser for behavioural categorisation of DTs, leveraging key Trust Evaluation Categories including safety, privacy, security, reliability, resilience, uncertainty & dependability, and ecosystem goal alignment. A DT simulator is developed to generate normal, unpredictable, and malicious DT behaviours, facilitating controlled experiments in a scalable DT ecosystem. Experimental results demonstrate that the TA achieves high accuracy, above 85% in detecting DT behavioural types, maintaining robust performance even as the ecosystem scales, with minor reductions attributable to network-induced delays and overlapping behaviour patterns. The proposed approach highlights the effectiveness of trust-based behavioural analysis for ensuring resilient, secure, and accountable operation in complex collaborative ecosystems.
Modern enterprise environments operate through interconnected workflows that change rapidly under shifting market trends, regulatory updates, and unpredictable workloads. Conventional automation systems rely on fixed logic, which often fails when faced with new patterns or sudden changes in business processes. Such rigidity causes frequent interruptions, slower task completion, and higher operational costs during critical transitions. Self-learning agentic Artificial intelligence (AI) cloud platforms introduce adaptive intelligence to address these challenges. These platforms consist of autonomous decision agents running on a shared cloud infrastructure. Each agent observes real-time operational data, detects emerging trends, and adjusts process strategies without stopping active tasks. Reinforcement learning methods allow agents to refine their decision-making through feedback-driven rewards, while transformer-based decision models help analyze complex data flows and uncover hidden dependencies between processes. The cloud foundation allows coordinated updates across all agents, creating a dynamic ecosystem that adapts continuously as business conditions evolve. The proposed system can be evaluated using several performance indicators. Key metrics include task completion speed, process accuracy, automation coverage, system response latency, and fault recovery duration. Energy usage and throughput under dynamic load can further reflect the platform's efficiency and stability. Experimental evaluations have reported faster convergence, reduced downtime, and higher accuracy compared to traditional static automation setups. Such outcomes demonstrate the ability of self-learning agentic AI platforms to build enterprise automation systems that grow in capability with continuous experience while supporting uninterrupted operations in complex digital environments.
