Cloud-native applications are increasingly vulnerable to high-traffic HTTP flood attacks, which can overwhelm scalability mechanisms and significantly degrade service availability. Traditional mitigation strategies focus on securing traffic filtering, detecting, and mitigating HTTP DDoS attacks. However, these strategies overlook architectural approaches and resilience through scalable mechanisms. This study presents an empirical evaluation of cloud-native application resilience and scalability under HTTP DDoS attacks, applying chaos engineering principles. We conducted extensive controlled HTTP flood attack experiments totalling over 80 hours within Amazon Web Services (AWS) infrastructure, generating more than 10 million HTTP requests. We systematically evaluated and compared four scaling and resilience mechanisms: Default scaling, Warm-up scaling, 2-Desired capacity scaling, and virtual private cloud (VPC)-based scaling. The results were analysed both technically and statistically. We leveraged native scaling and security mechanisms offered by cloud providers, including auto-scaling, application load balancers, security groups, and multi-availability zone deployment. Our findings demonstrate that while all mechanisms tend to have similar behaviour under low-load circumstances, they have different behaviour under high-intensity attacks. The VPC-based scalability mechanism demonstrates improved performance by reducing the failure rate by 12
Lung cancer is considered as the deadliest type of cancer, a good diagnosis as well as effective treatment is based on early detection. By utilizing Deep Learning (DL) techniques, the medical practitioner’s burden is reduced since these models can automate the diagnosis and disease classification. However, these models often suffer from modest accuracy and scalability issues, which affect the model’s disease detection performance. Consequently, a Federated Learning (FL)-based Global Data Trained Disease Detection (GDTDD) model, to provide better disease detection in hospitals. Initially, obtained local data (lung CT images) from different hospitals were pre-processed using a median filtering technique. Then these pre-processed images get segmented with the utilization of a multi-Convolutional layer-assisted U-Net (MCAU-Net) model. From these segmented images, features including Multi-texton, Median Binary Pattern (MBP) and Improved Local Gabor Transitional Pattern (LGTrP) features were extracted. Then these extracted features were fused utilizing multi-scale channel attention module (MS-CAM) based feature fusion technique. Then, trained the hybrid Recurrent Neural Network (RNN) and Modified Attention layer Assisted Bidirectional Long Short-Term Memory (MAA-Bi-LSTM) with this fused feature and obtained the local model as output. Afterwards, these local models were transmitted to Federated Learning, which merges and aggregates these locally trained models and provides a global trained model. This globally trained model is stored on the blockchain to ensure security and privacy. Then this global model is trained with fused features for accurate and effective disease detection. Moreover, this MAA-Bi-LSTM+RNN model’s effectiveness in disease detection is proven by the experimental findings. The Precision and F-measure values of the proposed model are 95.48
Efficient task scheduling and resource provisioning remain major challenges in cloud platforms due to workload heterogeneity, resource variability, and conflicting Quality of Service (QoS) objectives. This paper proposes an autonomic scheduling framework based on Grouped Moth-Flame Optimization (GMFO) to improve cloud task-to-resource mapping under dynamic workload conditions. The main methodological contribution is the adaptation of Moth-Flame Optimization through grouped population search, local subgroup leadership, inter-group knowledge sharing, and fission-fusion based diversity control. These mechanisms are designed to strengthen the exploration-exploitation balance and reduce premature convergence in large-scale scheduling spaces. The proposed framework formulates scheduling as a multi-objective optimization problem that considers makespan, execution cost, energy consumption, task acceptance, and load balancing. CloudSim-based experiments compare GMFO with established metaheuristic schedulers, including SMO, EGWO, PSO, BAT, and ABC. The results show that GMFO provides more stable and effective scheduling performance across the evaluated QoS metrics. Overall, the study demonstrates that grouping-driven MFO can provide a scalable and adaptive basis for autonomic cloud scheduling.
The major cloud providers offer their idle compute infrastructure through spot instances, discounted virtual machines that can be interrupted at any time. Although the range of spot instance types is broad, the possibilities expand even further due to factors such as variable discount rates, resource availability, and interruption probabilities, all of which vary by geographical region. In Amazon Web Services, spot-based applications, such as web services and batch processing, can take advantage of the EC2 Spot Service, which reduces operational complexity by providing mechanisms for fleet provisioning through allocation strategies. However, EC2 Spot Service for fleet provisioning has certain restrictions, which come with two key limitations: 1) the inability to query potential fleet prices without making actual reservations and 2) the restriction to a single predefined geographical region. This paper introduces a prediction-based approach for spot fleet provisioning based on monitoring across regions that preserves the capabilities of EC2 Spot Service while overcoming the mentioned constraints. Our solution removes the single-region restriction and predicts fleet prices before launch. We evaluate it on AWS across nine regions and two families of x86 instances, compute- and memory-optimized, for fleet sizes of up to 1,500 vCPUs. Our results show that this approach can exploit regional price differences of up to 80.5
With the rapid development of the Internet of Things (IoT) and edge computing, Federated Learning (FL) has emerged as a promising distributed framework capable of effectively leveraging distributed devices for machine learning tasks while preserving data privacy. However, in practical scenarios characterized by significant device heterogeneity and communication delays, traditional synchronous FL often suffers from inefficiency due to the straggler effect. Although Asynchronous Federated Learning (AFL) has been proposed to mitigate this, it still faces challenges regarding model staleness and training bias, which hinder convergence performance. To address these issues, this paper presents an asynchronous framework named FedQS. First, FedQS employs a multi-dimensional staleness evaluation mechanism that dynamically assesses updates by combining the similarity between local and global models with client latency metrics. Second, to resolve the coupling between training bias and model staleness, we implement a decoupling solution via a queue scheduling algorithm. This algorithm reprocesses high-staleness model parameters on selected faster helper clients using only the helpers’ local private data while preserving the provenance of the original uploading clients, thereby aiming to reduce stale-update effects in aggregation. Finally, during the aggregation phase, the framework recalibrates aggregation weights based on client contributions to reduce training bias and improve global model accuracy. Comparative experiments on Fashion-MNIST and CIFAR-10 datasets demonstrate that FedQS outperforms four evaluated representative baselines—FedAsync, FedBuff, KAFL, and CAFL—achieving an average accuracy improvement of 8.62
Container clusters are widely used to run containerized applications in cloud environments. A key feature of these clusters is autoscaling, the capability to automatically adapt computing resources in response to changing workloads. Autoscaling requires determining resource allocations for successive deployment windows. One strategy to address autoscaling is to use resource allocation algorithms that compute a new allocation for each deployment window based on workload information, but independently of previous allocations. We refer to this approach as history-unaware. A crucial aspect of history-unaware autoscaling is the transition between successive deployment windows, that is, the process of shifting the cluster from one allocation to the next. This transition is particularly challenging because the independence of allocations may result in substantial differences between them. In this paper, we present the Recycling-Based Transition (RBT) algorithm to tackle this challenge. The RBT algorithm introduces a method to identify suitable sets of containers and nodes to preserve across deployment windows, and then orchestrates the required operations to complete the transition. It provides the necessary complement to previous resource allocation approaches that did not address transitions, thereby establishing the foundation of a complete history-unaware autoscaling strategy. To evaluate the suitability of this strategy, we developed a basic simulator that implements history-unaware autoscaling as well as other strategies, including the horizontal autoscaling used in industry-standard systems. Experimental results show promising outcomes: history-unaware autoscaling achieves cost reductions in cluster deployments compared with industry-standard horizontal autoscaling. The impact of the RBT algorithm on cost savings, compared to a baseline algorithm, is also presented in the experimental section.
Cloud-edge collaborative inference, particularly the synergy between edge-deployed small models and cloud-based large models, demonstrates extensive application value in meeting the high-accuracy and low-latency demands of emerging artificial intelligence. However, existing dynamic offloading methods exhibit limitations in modeling the “many-to-many” high-order resource competition that arises when numerous small models concurrently access shared large model interfaces. They lack a unified representation capable of adapting to time-varying network congestion and often ignore the spatial exclusivity of concurrent tasks when sharing bandwidth and large model interfaces. To overcome these limitations, this paper proposes a spatiotemporal hypergraph-driven framework integrating high-order topological feature extraction with dynamic resource modeling. This framework introduces dynamic hypergraph sequences to naturally encompass local conflict domains, mitigating the topological blind spots and “over-smoothing” issues inherent in traditional pairwise graphs. In the first stage, a spatiotemporal hypergraph neural network (ST-HGNN) encodes complex spatial dependencies and predicts temporal resource bottlenecks. Subsequently, a weighted multi-dimensional hypergraph matching (WTHM) strategy transforms the NP-hard mixed-integer nonlinear programming problem into an efficient heuristic search. Trace-driven simulation results based on real-world urban trajectories (EUA and T-Drive) demonstrate that the proposed framework achieves a robust normalized VCI score of 0.64 even during extreme congestion peaks. By consistently achieving an absolute VCI improvement of 0.17 to 0.38 over traditional bipartite-graph and proximity-based baselines, the framework significantly mitigates cascading queuing failures. Moreover, the developed dual-timescale asynchronous architecture strictly confines online execution to the millisecond scale, validating its capability to balance inference accuracy and latency across dynamic cloud-edge environments.
Preemptible cloud services utilize idle resources from contract-based services. The resources allocated to these low-priority services can be preempted by contract-based services during periods of overload, so cloud platforms offer low prices to attract cost-sensitive users. While preemptible cloud services enhance resource utilization and boost revenue, excessive admission of service requests may induce overbooking effects, resulting in an excessively high preemption rate that undermines user satisfaction and ultimately lowers revenue. Conversely, insufficient admissions may result in underutilized resources. The stochastic nature of idle resource capacities necessitates real-time dynamic optimization of admission control. The primary difficulty in optimizing dynamic admission control lies in the fact that fluctuations in idle resource capacity are a special category of stochastic fluctuations characterized by non-stationarity, where the stochastic distribution patterns evolve over time. To address this difficulty, we first formulate the admission control for preemptible cloud services with overbooking effects as a Non-Stationary Markov Decision Process (NSMDP). Then, an improved soft actor-critic (SAC) framework is proposed for the NSMDP model. In this framework, the regularization and feature-enhanced LSTM algorithm is employed for turning point detection (TPD) in non-stationary stochastic fluctuation patterns. On this basis, a soft actor-critic algorithm based on turning point detection (TPD-SAC) is developed to achieve real-time resolution of NSMDP. Experiments with Google Cloud datasets validate the effectiveness of the proposed algorithm in non-stationary environments. Furthermore, simulation experiments were conducted to analyze the revenue contributions of dynamic admission strategies under two pricing mechanisms. Experimental results indicate that compared to strategies under the uniform discount mechanism, dynamic admission strategies achieve higher revenue contributions under the interruption-based discount mechanism.
Epilepsy is a common and serious neurological disorder affecting approximately 50 million people worldwide. Recent advancements in artificial intelligence have facilitated significant improvements in epilepsy prediction methodologies. Research has increasingly focused on developing prediction models that utilize EEG signals and deep learning, resulting in a notable increase in published findings. Nonetheless, existing research faces two critical challenges. First, prevalent model architectures rely on traditional neural networks trained in centralized frameworks, which often fail to comply with privacy regulations. Second, current deep learning strategies, including Graph Neural Networks, utilize static graph models that overlook the dynamic nature of EEG signal networks, thereby failing to capture complex nonlinear correlations within brain functional networks. This paper introduces a patient-dependent, privacy-preserving Federated Learning framework that incorporates DygonNet, a prediction model based on a spatiotemporal dynamic graph neural network alongside a localized learning model at each federated client within a cloud-based simulation. The CHB-MIT, SWEC-ETHZ, and TJU-HH iEEG datasets are utilized as independent federated clients, each conducting federated training on secure EEG data. Innovatively, a central cloud server on Google Colab Pro aggregates model updates via Federated Averaging and FedProx algorithms without accessing raw patient data. The model delineates the dynamic graph framework of EEG signals and integrates the Transformer model with a dynamic graph neural network for a comprehensive analysis of EEG signal characteristics, proposing a hierarchical graph pooling mechanism driven by attention within a Federated Learning context. Experimental results indicate that the model demonstrates superior epilepsy prediction efficacy across both public and private datasets.
PM2.5 monitoring is an important task for air-quality management and healthcare-oriented smart environments, especially in cities where fine particulate pollution may affect respiratory and cardiovascular health. In IoT-enabled monitoring systems, PM2.5 data are usually collected by distributed sensors. However, the collected records are often incomplete because of sensor faults, low battery power, unstable communication links, or temporary device maintenance. These missing values may weaken subsequent air-quality assessment and reduce the reliability of health-related decision support. Therefore, it is necessary to develop an efficient method for estimating missing PM2.5 records, particularly when the available monitoring data are sparse. To address this problem, this paper proposes PMLSH, a locality-sensitive hashing based PM2.5 prediction method. The key idea is to convert high-dimensional city-time PM2.5 records into compact hash indices, so that similar entries can be retrieved without exhaustive pairwise comparison. The missing PM2.5 value is then estimated from the known values of its neighboring entries. In this paper, “PM2.5 prediction” specifically refers to estimating missing values within a partially observed monitoring matrix, rather than forecasting PM2.5 concentrations over a future time horizon. Since LSH supports efficient approximate similarity search, PMLSH reduces the time cost of missing-value prediction while maintaining a low prediction error. Experiments on a real PM2.5 dataset collected from an official environmental monitoring source show that the proposed method achieves lower prediction error and less computational time than the compared methods.
Task offloading in vehicular networks (VNets) is complicated by fluctuating channels, dynamic topologies, and bursty task arrivals. No single algorithmic paradigm performs reliably across all dynamic conditions, making hybridization a compelling approach. Based on patterns observed in the literature, this survey classifies hybrid algorithms into six basic categories and two extended categories. We characterize environmental dynamism along three dimensions—channel dynamics, topology dynamics, and task dynamics—and define four operating regimes from static to highly dynamic. We review studies across these categories and regimes and compare their hybridization mechanisms, validation conditions, and reported limitations. Based on this analysis, we identify key research gaps in online dynamism detection, cross-category benchmarking, and runtime meta-control. Unlike prior surveys that primarily catalog algorithm combinations, this survey provides a structured taxonomy, an environmental characterization framework, and a diagnostic perspective to guide future research.
Cloud computing has transformed the modern IT infrastructure with its scalable, cost-effective, and on-demand services. Distributed Denial-of-Service (DDoS) attacks, however, are growing in number and sophistication, and they can adversely affect service availability, negatively impact performance, cause financial losses, and damage reputation, particularly in the cloud environment. Conventional detection methods are not always adequate to deal with the dynamic and massive nature of these attacks. To solve these challenges, this paper proposes a machine learning-based Resource-Optimized DDoS Attack Detection (RO-DAD) framework that enhances detection efficiency through a multi-stage feature selection pipeline. The proposed framework integrates Domain Knowledge (DK) filtering, Least Significant (LS) feature elimination, and Most Significant (MS) selection, where the MS stage is implemented using an embedded tree-based feature selection mechanism. This design enables systematic identification of the most informative features while reducing computational complexity and eliminating redundant attributes. The framework is evaluated using the CICDDoS2019 dataset, and experimental findings show that the proposed approach achieves high detection performance, reaching up to 99
In today’s interconnected world, IoT devices have transformed daily life by automating processes and improving productivity. However, as these devices handle large volumes of sensitive data, robust security mechanisms are essential. Traditional encryption approaches face growing challenges with the advent of quantum computing, necessitating the development of quantum-resistant solutions. This study presents a future-proof approach for securing IoT data using multi-key homomorphic encryption, proxy re-encryption, and a lattice-based key-switching mechanism based on the Learning With Errors assumption. The proposed scheme is evaluated both theoretically and experimentally on a dataset of 1000 IoT devices. Under the chosen parameter settings, the scheme achieves a re-encryption time of approximately 140 ms, an encryption/decryption time of around 280 ms, and a key generation time of 80 ms. The results demonstrate that the proposed approach offers competitive performance in terms of key generation, encryption/decryption, and re-encryption time, while showing lower runtime compared to selected baseline schemes in our experimental setup.
Over the last few years, we have experienced a growing avalanche of information generated by devices connected to the Internet derived from the widespread acceptance of technology in society. These devices have enabled the development of IoT applications that facilitate and improve the quality of life of people in different areas such as industry or healthcare. As applications advance, their complexity increases due to the interconnection with more devices and the data they generate. This places greater demands on the architecture responsible for transferring and processing information, which can have a negative impact on performance. This implies that Quality of Service (QoS) has to be managed with precision. Therefore, it is essential to select a suitable architecture for data processing and analysis that meets QoS requirements such as response time or latency. In this paper, we present a set of parameters for evaluating and comparing analytics performance between two architecture alternatives: Cloud Computing and the emerging paradigm of the Computing continuum. These parameters influence QoS in evaluating and comparing these architectures; as validation, we provide a case study where the suitability of each architecture can be observed and detected depending on the application context. Based on the empirical analysis, we find that Cloud computing excels in environments with smaller data sets and limited devices; conversely, the Computing continuum proves superior in scenarios with larger data sets and numerous end devices.
As an essential security primitive, searchable encryption safeguards the security and availability of cloud-stored data. Attribute-Based Searchable Encryption (ABSE) achieves fine-grained access control while conducting retrieval based on keywords. Nevertheless, most existing ABSE schemes neglect privacy leakage incurred by misbehaviors of data owners, and the retrieved results fail to support verification. In practical scenarios, data owners may intentionally or unintentionally disclose secret keys and secret values of encrypted indices to third parties, resulting in the leakage of file contents and search keywords. Meanwhile, to reduce communication and storage overhead, cloud servers may return incomplete search results to users. To tackle these issues, this paper proposes a searchable encryption scheme that can resist privacy leakage from data owners and verify returned search results. This scheme incorporates sanitization operations to guarantee file security, user search keyword security, and integrity verification for results returned by the cloud server. We prove the security of the scheme, including indistinguishability against selective ciphertext-policy and chosen-plaintext attack (IND-sCP-CPA), as well as the unforgeability of index and trapdoor. Moreover, the performance analysis validates the feasibility of the proposed scheme.
Adversarial perturbation represents an acute problem inadequately covered by current defenses that have not sufficiently secured Machine Learning-based Intrusion Detection System (IDS) running in cloud environments against evasion attacks. Namely, adversarial manipulations can be easily carried out with minimal changes in input data, resulting in failed detection, false-positive classification of attacks as benign traffic, and model corruption without raising traditional alerts. Existing defensive techniques further aggravate the problem, since centralized monolithic models demonstrate vulnerability to adversarial evasion, federated implementations open up data poisoning vulnerabilities due to distributed nature of training, and black-box solutions do not allow explaining decisions in terms comprehensible to humans. In order to tackle this trifecta of challenges, we suggest an innovative Adversarial-Resilient Federated Ensemble Learning (AR-FEL) system. Our solution unites three key ideas: (i) an adversarially trained meta-classifier ensemble consisting of four different Machine Learning classifiers – XGBoost, Random Forest, Gradient Boosting, and Multi-Layer Perceptron (MLP), aggregated into a stacked logistic regression, capable of counteracting both decision boundary adversarial attacks and semantics-based manipulation aimed at deceiving monolithic models; (ii) a linguistic-model-driven post-hack analysis module responsible for interpreting numeric anomaly scores and translating them into human-readable, actionable attack scenarios – an approach that allows overcoming limitations of purely numeric explanations like SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME); and (iii) using federated learning, which not only ensures confidentiality of training but also serves as a tool for preventing local data poisoning from being propagated to the global model. The effectiveness of our proposed model was tested against adversarial attacks such as Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and Carlini Wagner (adversarial attack), as well as adversarial perturbation aimed at manipulating model’s decision boundary and semantics. In our experiments, the performance metrics of Adversarial-Resilient Federated Ensemble Learning (AR-FEL) were 99.91
Complex nonlinear problems are typically solved by optimization methods, and metaheuristic algorithms have attracted significant attention because of their ability to solve complex optimization problems. Most of the available approaches, however, have weaknesses, including premature convergence, a lack of balance between exploration and exploitation and slow convergence within high-dimensional spaces. To overcome these shortcomings, in this paper, the hybridization of the sine cosine algorithm (SCA) and butterfly optimization algorithm (BOA) and the hybrid sine cosine-butterfly optimization algorithm (SCA-BOA), which have the global exploration power of the sine cosine algorithm and the high local exploitation power of the butterfly optimization algorithm, are suggested. The suggested solution works in two phases. In the first step, the sine cosine mechanism improves the diversity in the population and conducts a global search with sinusoidal operators. At a later stage, the butterfly optimization strategy increases the search near promising regions, enhancing local refinement. This hybrid design permits a more balanced trade-off between exploration and exploitation as well as improved convergence stability. The proposed algorithm was evaluated using classical benchmark functions and the IEEE CEC 2022 benchmark suite under identical experimental settings. Ten standard benchmark functions were tested using the algorithm and compared to nine existing metaheuristic algorithms under the same experimental conditions. The findings show that the SCA-BOA has better performance in terms of fitness values, convergence speed, and stability of the solution. The experimental results demonstrate that the proposed hybrid SCA-BOA achieved the best overall average ranking score (2.3) among the compared optimization algorithms and produced competitive mean fitness and standard deviation values across diverse benchmark landscapes. Furthermore, the proposed algorithm was validated using the IEEE CEC 2022 benchmark suite, where convergence analysis demonstrated improved convergence stability and adaptive search behavior on unimodal, multimodal, hybrid, and composition benchmark functions. In addition, population diversity analysis quantitatively confirmed the balanced exploration–exploitation capability of the proposed hybrid algorithm and its effectiveness in avoiding premature convergence during optimization. These results demonstrate the effectiveness of the proposed algorithm and its ability to address complex continuous optimization problems in the field.
A cloud-native distributed pipeline integrating statistical synthetic data generation with adaptive routing and infrastructure optimisation is proposed for unmanned aerial vehicle (UAV)-based emergency medical logistics in data-sparse rural environments. The synthetic data generation component employs a kernel density estimation (KDE)-based statistical model that learns mission attribute distributions from real operational data without requiring neural generative architectures, making it well-suited for data-sparse deployment contexts. The framework combines a three-layer edge-cloud architecture with stochastic demand augmentation, enabling scalable mission orchestration and real-time reconfiguration across distributed UAV networks. A dataset of 611 real-world missions was expanded to 3,055 simulated events through synthetic generation achieving a distributional deviation of 0.0093. Evaluated in simulation across 3,000 routing instances, the framework achieves a mission success rate of 98.47
Cloud computing environments are increasingly vulnerable to sophisticated cyber threats due to the rapid growth of cloud-native infrastructures, high-volume network traffic, and evolving attack strategies such as zero-day exploits, adversarial attacks, and polymorphic malware. Traditional intrusion detection systems often struggle to maintain high detection accuracy, scalability, adversarial robustness, and real-time processing capability in dynamic cloud environments. To address these challenges, this study proposes an adaptive Intrusion Detection System-as-a-Service (IDSaaS) framework using a Modified Attention-Gate based BiLSTM-GRU architecture integrated with GAN-assisted synthetic anomaly generation. The proposed framework combines Bidirectional Long Short-Term Memory networks for temporal dependency learning, Gated Recurrent Units for computational efficiency, a custom attention mechanism for dynamic feature prioritization, a modified recurring gate structure for optimized information retention, and a Revamping Input Sequence Weighing Structure (RISWS) loss function for improved anomaly classification performance. Experiments were conducted using the CICIDS-2018, UNSW-NB15, SCAPY-based, and SCADANet datasets within an AWS cloud deployment environment. The proposed model achieved strong intrusion detection performance with 97.91
The rapid deployment of 5G networks has introduced significant challenges in managing scarce and valuable resources, such as core network cache memory. Ensuring efficient and fair allocation of these resources across heterogeneous service slices is essential for maintaining high Quality of Experience (QoE) and guaranteeing system-wide stability. This paper presents a novel service-aware cache allocation framework for 5G network slicing that leverages cooperative game theory to achieve fairness and efficiency. The proposed approach operates in three phases: (i) an admission control and prioritization mechanism that dynamically ranks slices based on critical Quality of Service (QoS) metrics—latency, throughput, reliability, and availability; (ii) an initial proportional allocation that respects guaranteed minimum resources while aligning with priority scores; and (iii) a cooperative bankruptcy game model, where the Shapley value is applied to redistribute cache resources under demand overload conditions equitably. Extensive simulations demonstrate that the Proposed Method outperforms conventional proportional fairness (PF) and max-min fairness (MMF) schemes, achieving up to 3
