Distributed Machine Learning PDF

Distributed Machine Learning VISHNUVARDHAN SOMPALLI RAJA SHEKHAR DASARI MALLISETTI MANOJ [email protected] [email protected] [email protected] SUCHAN CHOWDAR...

Distributed Machine Learning VISHNUVARDHAN SOMPALLI RAJA SHEKHAR DASARI MALLISETTI MANOJ [email protected] [email protected] [email protected] SUCHAN CHOWDARY ADUSUMALLI PREETHI REDDY BOOSI [email protected] [email protected] Abstract—A crucial paradigm for training large-scale models II. L ITERATURE S URVEY on enormous datasets over numerous workstations is distributed machine learning, or DML. DML improves model training effi- ciency and scalability by distributing computing duties, allowing A. Power Allocation Schemes Based on Deep Learning for for faster convergence and handling of data that is larger than Distributed Antenna Systems what can be handled by a single machine. This article examines G. Qian, Z. Li, C. He, X. Li and X. Ding different DML implementation frameworks and strategies with an emphasis on fault tolerance, scalability, and communication 1) Introduction: The exponential rise in data transmission effectiveness. Using a well-known framework (such as TensorFlow requirements in cellular networks, especially with the introduc- or PyTorch), we describe an implementation of a distributed training algorithm and compare its effectiveness with conven- tion of 5G technology, requires Distributed Antenna Systems tional centralized techniques. Our results show notable gains in (DAS) to allocate power efficiently in order to maximize model accuracy and training time, underscoring the benefits of energy and spectrum efficiency. Real-time applications cannot using distributed computing resources for machine learning tasks. benefit from the high processing complexity of traditional iterative power allocation algorithms. 2) Methodology: In order to approach the conventional sub- I. I NTRODUCTION gradient algorithm for power allocation in DAS, this study suggests using a Deep Neural Network (DNN) model. Our The science of machine learning has undergone a revolution goal is to teach the DNN a nonlinear mapping that allows for with the introduction of big data, which has made it possible effective real-time power allocation by using a dataset that is to create increasingly complex models that can learn from derived from channel realizations and optimal power allocation enormous volumes of data. However, the processing power schemes produced by the sub-gradient technique. and computational resources of classic machine learning tech- niques are frequently limited, particularly when working with 3) Key Results and Their Significance: According to the large datasets. Distributed machine learning (DML), which simulation results, the DNN approach greatly reduces online enables the training of models over numerous computers or processing time—by at least three orders of magnitude—while network nodes, has gained popularity as a workable solution achieving over 92% accuracy when compared to the conven- to these problems. tional sub-gradient methodology. This illustrates how DNNs DML helps handle datasets that are too large to fit in a can offer near-optimal performance at a reduced complexity, single machine’s memory while simultaneously speeding up enabling real-time wireless communication system applica- the training process. DML improves scalability and efficiency tions. by utilizing the combined processing capacity of numerous 4) Discussion: The results imply that operational efficiency machines by dividing data and calculations across different in DAS can be improved by including machine learning nodes. Moreover, DML is now more feasible and available techniques, such as DNNs, into power allocation strategies. to a wider range of scholars and practitioners due to recent However, the study may not fully depict dynamic real-world developments in communication protocols and frameworks. settings due to its reliance on perfect channel state information. The purpose of this research is to investigate the frameworks Subsequent investigations may tackle these constraints by and methods used in distributed machine learning. We will investigating more resilient machine learning methodologies examine DML’s drawbacks, such as communication overhead, and adaptive real-time systems. fault tolerance, and the difficulty of distributed nodes syn- 5) Conclusion: This work effectively demonstrates the use chronizing with one another. We will also showcase a case of a DNN for power allocation in DAS, obtaining performance study that compares the performance of a distributed training that is on par with conventional techniques while requiring algorithm to traditional centralized training techniques. Our a significant decrease in computation. Subsequent research goal in doing this study is to demonstrate how DML might endeavors will center on optimizing these models and inves- hasten the creation and implementation of machine learning tigating substitute machine learning methodologies to further models in practical settings. augment efficacy and versatility in wireless systems. B. Self-Organizing Democratized Learning: Toward Large- 2) Methodology: This publication provides academics who Scale Distributed Learning Systems are interested in using ML to solve networking problems Minh N.H. Nguyen, Shashi Raj Pandey, Tri Nguyen Dang, Eui- with a useful guide. It goes over important machine learning Nam Huh, Nguyen H. Tran, Walid Saad, Choong Seon Hong (ML) ideas, investigates deep learning architectures that are 1) Introduction: New cross-device AI applications require pertinent to networking, finds tools and datasets that are freely a move away from centralized learning systems and toward accessible, and talks about ways to maximize ML efficiency, large-scale distributed AI frameworks that can handle per- such as Split Learning and Federated Learning. sonalization and data protection issues while supporting 3) Key Findings: The results highlight the reciprocal inter- collaborative learning. Present-day federated learning (FL) action between machine learning and networking, classifying techniques exhibit inadequate model performance in real- methods as networks serving machine learning (N4ML) and world situations due to their inability to strike a compromise ML serving networks (ML4N). This framework facilitates the between generalization and personalization. creation of innovative ML-powered networking solutions by 2) Methodology: The Dem-AI philosophy, which stresses highlighting current research gaps and offering researchers hierarchical self-organization among learning agents, served practical examples and tools. as the inspiration for the new distributed learning algorithm, 4) Discussion: The study highlights several potential appli- DemLearn, which is presented in this work. To improve cations of machine learning (ML) and networking, but it also model performance, the method formulates recursive hierar- highlights drawbacks, such as the lack of high-quality training chical generalized learning problems and uses agglomerative data and the requirement for interpretability in ML models. clustering to organize agents according to learning features. Given the complexity of the industry, resolving these issues 3) Key Findings: Empirical findings on industry standard is essential to the wider use of ML solutions in networking datasets such as MNIST and CIFAR-10 show that DemLearn situations. outperforms traditional FL methods by considerably enhancing 5) Conclusion: This paper highlights the possibilities for the generalization performance of client models while retain- further research and innovation by providing a thorough ing robust specialization capabilities. This research shows how overview and useful insights into the nexus of machine learn- democratized learning systems can successfully manage the ing and computer networking. It seeks to support continued competing demands of personalization and generalization in investigation of novel techniques and datasets while also distributed settings. promoting improvements in machine learning applications in 4) Discussion: While the Dem-AI framework presents networking by helping researchers through current tools and useful insights for creating customized AI systems, it also difficulties. highlights issues with how well models adapt to changing conditions and how difficult it is to handle several learning D. Data Poison Detection Schemes for Distributed Machine tasks. To overcome these constraints, future studies should Learning concentrate on improving robustness, diversity, and the incor- Y. Chen, Y. Mao, H. Liang, S. Yu, Y. Wei and S. Leng poration of cutting-edge transfer learning strategies. 5) Conclusion: DemLearn successfully combines general- 1) Introduction: By using several nodes, distributed ma- ization and specialization in a hierarchical framework, which chine learning (DML) makes it possible to process large is a significant achievement in distributed machine learning. datasets, but this added complexity also makes it more vul- Subsequent research endeavors will delve into the expansion nerable to data poisoning assaults. The integrity of distributed of these concepts to multi-task learning and provide additional systems can be seriously jeopardized by these attacks, evidence of their efficacy in practical settings, ultimately emphasizing the essential need for efficient data poisoning augmenting the potential of customized intelligent systems. detection systems. 2) Methodology: Two different data poisoning detection C. Machine Learning With Computer Networks: Techniques, strategies that are suited for basic- and semi-DML scenarios Datasets, and Models are presented in this research. The basic-DML technique uses H. Afifi, S. Pochaba, A. Boltres, D. Laniewski, J. Haberer, L. a mathematical model to calculate the ideal number of training Paeleke, R. Poorzare, D. Stolpmann, N. Wehner, A. Redder, E. loops and a cross-learning data assignment mechanism to Samikwa, and M. Seufert. detect tainted data. On the other hand, the semi-DML strategy 1) Introduction: Growing expectations for network perfor- improves learning protection by optimizing resource allocation mance measures like latency and dependability are the and enhancing data poisoning detection. result of computer networks becoming more complicated and 3) Key Findings: According to simulation results, the sug- larger than they were a decade ago. Traditional computational gested detection strategy can improve model accuracy by as methods have not kept up with this trend. Machine learning much as 60% for logistic regression and up to 20% for support (ML) has shown promise in a variety of domains, including vector machines in basic-DML. Furthermore, the semi-DML networking, where it can improve performance and decision- technique may minimize resource waste by 20–100% with making. However, issues with data quality, resource limita- appropriate resource allocation, improving the effectiveness tions, and model interpretability still need to be addressed. and economy of the detection process.  4) Discussion: Although the suggested strategies demon- F. Managing Distributed Machine Learning Lifecycle for strate encouraging outcomes in terms of increasing resource Healthcare Data in the Cloud efficiency and model correctness, there are obstacles associ- E. Zeydan, S. S. Arslan and M. Liyanage ated with their application in dynamic contexts with fluctuating 1) Introduction: The healthcare industry faces difficulties attack intensities. In order to guarantee that detection methods with data management, cost, and integrating cutting-edge tech- continue to function well in a variety of scenarios, future re- nologies to enhance patient care. The generation of sensitive search must examine the trade-off between increased security biological data in huge volumes might compromise security and resource usage. and privacy when handled using traditional approaches. This 5) Conclusion: This work greatly enhances the security emphasizes the need for strong artificial intelligence (AI) of distributed machine learning systems by offering a thor- and machine learning (ML) frameworks that can handle the ough methodology for identifying data poisoning in DML data efficiently throughout its lifecycle. systems. The research establishes the foundation for upcoming 2) Methodology: The end-to-end data engineering pipeline improvements in protecting machine learning models against proposed in this research is intended for effective ML life- adversarial assaults while maximizing resource utilization by cycle management in the healthcare industry. The strategy providing workable detection algorithms for both basic and incorporates federated learning (FL) and other distributed semi-DML. computing approaches to protect data privacy and promote E. Bayesian Optimization-Driven Adversarial Poisoning At- inter-institutional collaboration without disclosing private in- tacks Against Distributed Learning formation. The architecture includes modules designed to M. Aristodemou, X. Liu, S. Lambotharan and B. AsSadhan handle specific difficulties in healthcare data management, including data source integration, ingestion, analysis, storage, 1) Introduction: The enormous volumes of sensitive data and visualization. used in the Metaverse’s applications across numerous indus- 3) Key Findings: The results show that utilizing cloud in- tries present serious privacy concerns, despite the fact that it is frastructures and AI/ML frameworks greatly improves health- envisioned as a revolutionary, immersive digital environment. care solutions by increasing scalability, security, and efficiency. While federated learning (FL) and split learning (SL), two Notably, the application of FL fosters a collaborative training popular distributed learning frameworks, attempt to reduce pri- paradigm and facilitates compliance with strict healthcare vacy hazards, they are still susceptible to adversarial poisoning laws, which in turn improves patient outcomes and offers more assaults, therefore strong defenses are needed. customized treatment alternatives. 2) Methodology: This paper simulates adversarial behavior 4) Discussion: Healthcare data management could be revo- against FL and SL by developing two new poisoning attack lutionized by the suggested paradigm, but there are still major techniques based on Bayesian optimization. The suggested obstacles to overcome, including data quality, communication approaches, BO-FLPA and BO-SLPA, target the hidden layers overhead, and interoperability. Future studies should concen- of neural networks by relating prediction uncertainty with trate on addressing these obstacles to guarantee the successful optimization parameters using a mapping function. Standard and long-lasting application of AI/ML technologies in actual datasets like MNIST and CIFAR10 are used to test the system healthcare environments. and determine how various assaults affect model performance. 5) Conclusion: A viable avenue to improving the handling 3) Key Findings: The findings show that in both FL and SL of sensitive medical data is the integration of modern data scenarios, the suggested poisoning attacks seriously impair the engineering methods with AI/ML frameworks. Through the accuracy of global models, resulting in decreased prediction exploration of future research avenues and the resolution of confidence and higher output variances. This emphasizes how present difficulties, this strategy has the potential to enhance vital it is to have strong defenses against hostile activity in healthcare outcomes and optimize the overall efficiency of the privacy-sensitive Metaverse applications. healthcare system. 4) Discussion: Although the study successfully illustrates how Bayesian optimization may be used to create adversarial G. Extremely Randomized Trees With Privacy Preservation for attacks, its main emphasis on proof of concept restricts the Distributed Structured Health Data investigation of useful applications in real-world Metaverse A. Aminifar, M. Shokri, F. Rabbi, V. K. I. Pun and Y. Lamo scenarios. To further grasp the relevance of these findings, 1) Introduction: Incorporating machine learning (ML) and future research should focus on applying them to dynamic artificial intelligence (AI) into healthcare could greatly im- contexts like social networks and eye tracking. prove patient outcomes and decision-making. But maintain- 5) Conclusion: This study highlights the need for sophisti- ing patient privacy while analyzing dispersed medical data cated adversarial detection and defense techniques by offering continues to be a major difficulty, especially when data fundamental insights into the weaknesses of distributed learn- is dispersed across several sources like hospitals and per- ing frameworks in the setting of the Metaverse. The study sonal devices. The necessity for efficient privacy-preserving emphasizes the significance of protecting privacy and integrity strategies that enable the use of dispersed health data without in the developing field of immersive digital technologies by jeopardizing private patient information is discussed in this examining the effects of poisoning attacks. study.  2) Methodology: With privacy preservation in mind, we computations, the ADMM-type DITML performs better than provide k-PPD-ERT, a distributed extremely randomized trees the diffusion-type DITML in most cases, indicating the effi- technique specifically built for learning from decentralized cacy of the distributed frameworks while preserving competi- health data. When this algorithm is used on a cloud platform tive performance levels. like Amazon AWS, it may be extensively evaluated using a 4) Discussion: The suggested frameworks provide a work- variety of healthcare datasets, such as those relating to heart able way to learn metrics in dispersed contexts, which can disease, breast cancer, and mental health. greatly improve data usage in decentralized networks. How- 3) Key Findings: For the mental health datasets, our method ever, in very resource-constrained circumstances, the scalabil- produced up to 12.9% better F1-scores and noteworthy im- ity of the ADMM-type solution may be limited due to its provements in accuracy (ACC) and Matthews correlation co- requirement for additional processing resources. Subsequent efficient (MCC), demonstrating significant performance gains investigations may delve deeper into enhancing computing over current state-of-the-art distributed tree-based models. efficiency. These findings suggest that the k-PPD-ERT algorithm con- 5) Conclusion: The paper successfully addresses the issues tributes to more trustworthy healthcare analytics by maintain- of decentralized data and unlabeled pairings by presenting two ing both good predictive performance and patient privacy. efficient distributed semi-supervised metric learning frame- 4) Discussion: Although the suggested framework success- works. Through the use of these techniques, practitioners can fully strikes a compromise between privacy and data utility, it obtain strong metric learning skills without the requirement for relies on an honest but skeptical security model, which might centralized data, opening the door to more useful and scalable not apply in all real-world situations. In order to provide even applications across a range of industries. greater data protection and integrity, future work will need to investigate robust techniques to handle potentially harmful I. Machine Learning With Big Data: Challenges and Ap- actions among parties. proaches 5) Conclusion: High-quality machine learning models can A. L’Heureux, K. Grolinger, H. F. Elyamany and M. A. M. be developed by utilizing dispersed health data in a safe and Capretz effective manner, thanks to the k-PPD-ERT algorithm. This development is anticipated to boost patient care outcomes 1) Introduction: While the Big Data revolution holds great and healthcare analytics, providing a foundation for future potential to improve decision-making and insight discovery, investigations into more secure frameworks for data sharing. conventional machine learning techniques suffer from serious drawbacks from antiquated presumptions, like the need for H. Distributed Semi-Supervised Metric Learning the complete dataset to fit into memory. These difficulties are P. Shen, X. Du and C. Li caused by the distinct qualities of Big Data, such as its volume, 1) Introduction: More recently, algorithms for automati- velocity, variety, and authenticity, which make it difficult to cally learning measures from data constrained by similarity use data analytics effectively. The purpose of this study is to and dissimilarity pairs have been developed, representing a offer insights into appropriate machine learning approaches significant advancement in pairwise-constraint-based metric and to methodically handle these problems. learning. But current approaches usually presuppose central- 2) Methodology: This paper gathers and arranges Big Data- ized data gathering, which is not feasible for real-world scenar- related machine learning problems by classifying them based ios where data is dispersed over several nodes and frequently on the ”V” dimensions that give rise to these problems. It also comprises a large number of unlabeled pairings. The proposed examines recently developed machine learning techniques and distributed semi-supervised metric learning frameworks in this talks about how well they work to solve the problems that have paper can take advantage of both labeled and unlabeled data been identified. To show the connections between problems pairs in order to tackle these difficulties. and suitable machine learning solutions, a thorough matrix is 2) Methodology: The diffusion-type DSSML and the created. ADMM-type DSSML are two distributed semi-supervised 3) Key Findings: In the context of Big Data, the study metric learning frameworks that we provide. The ADMM- successfully classifies machine learning challenges and cre- type makes use of distributed ADMM optimization techniques, ates cause-and-effect relationships for each problem. It gives whilst the diffusion-type uses a ”adapt-then-combine” iterative practitioners a better grasp of how to overcome these barriers solution. Both frameworks build upon SERAPH, a centralized by connecting particular difficulties to machine learning tech- semi-supervised metric learning technique that enables effec- niques. This helps practitioners choose the best solutions for tive learning without requiring the centralization of original their particular use cases. data—only transmitting intermediate estimates during the pro- 4) Discussion: The results point to important areas for fu- cess. ture research as well as chances to improve current approaches 3) Key Findings: Simulation findings show that in most or create new machine learning paradigms to better address cases, the metrics obtained by using the suggested distributed open problems. Although the matrix is a helpful tool, its methods nearly match those obtained by using the centralized relevance may differ based on particular circumstances and SERAPH technique. Notably, even with more sophisticated the ever changing landscape of big data technology.  5) Conclusion: This paper provides a thorough overview K. Machine Learning and Deep Learning Techniques for of the difficulties encountered in machine learning with big Distributed Denial of Service Anomaly Detection in Software data, providing a solid understanding and an extensive matrix Defined Networks—Current Research Solutions to help practitioners and researchers alike. It establishes the N. S. Musa, N. M. Mirza, S. H. Rafique, A. M. Abdallah and foundation for upcoming developments in the field, with the T. Murugan goal of enhancing the integration of machine learning tech- 1) Introduction: While the introduction of Software De- niques in the Big Data environment, by highlighting crucial fined Networks (SDNs) has improved network adminis- research possibilities. tration, it has also made networks more vulnerable to security threats, especially Distributed Denial of Service (DDoS) as- saults. In order to increase SDN security against these threats, this paper emphasizes the urgent need for efficient detection J. RLPTO: A Reinforcement Learning-Based Performance- and mitigation mechanisms, with a focus on sophisticated Time Optimized Task and Resource Scheduling Mechanism for Machine Learning (ML) and Deep Learning (DL) techniques. Distributed Machine Learning 2) Methodology: This systematic review analyzes different X. Lu, C. Liu, S. Zhu, Y. Mao, P. Lio and P. Hui approaches, including ensemble learning, supervised learning, and unsupervised learning, and classifies current DDoS detec- 1) Introduction: Deep learning models are becoming more tion tactics in SDNs into ML and DL approaches. By utilizing and more complicated, requiring large amounts of processing these techniques, the suggested systems evaluate network power and time for training. This has made resource allocation traffic data and apply automated detection and mitigation in distributed computing clusters imperative. In order to protocols to improve security. maximize throughput and adhere to privacy regulations, IT 3) Key Findings: The examination of the suggested meth- organizations must establish strong scheduling techniques as ods shows that they are highly accurate at both identifying they use distributed systems to boost model performance. and thwarting DDoS attacks; some models even manage to detect attacks with 100% accuracy and very low false-positive 2) Methodology: This work uses deep reinforcement learn- rates. These results highlight how ML and DL approaches may ing for resource scheduling in distributed learning clusters, effectively defend against network threats and greatly increase effectively predicting future task demands by applying a task the resilience of SDNs. volume prediction technique known as LSP. Additionally, the 4) Discussion: Though the systems under examination study develops a scheduling technique based on reinforcement exhibit encouraging outcomes, fundamental constraints like learning, creating a reward function, action space, and state the requirement for practical validation and the creation of space that are specifically designed to manage computer extensive datasets continue to be crucial. Subsequent investi- resources in an adaptive manner according to task character- gations ought to delve into the integration of these systems istics. with extant security protocols and enhance their efficacy to 3) Key Findings: According to experimental results, the accommodate distinct forms of assault and heterogeneous RLPTO resource scheduling method outperforms conventional network environments. scheduling techniques like FCFS and FS, greatly increasing 5) Conclusion: This review emphasizes the critical role of the scalability and efficiency of distributed computing clusters. advanced ML and DL techniques in enhancing DDoS detection The results show that allocating resources according to the size and mitigation within SDNs, showcasing their potential for of the data and the beta value can maximize model accuracy improving network security. Continued research is essential to and minimize training time, with practical implications for address existing limitations and develop more sophisticated, handling large-scale machine learning jobs. scalable solutions for real-world applications. 4) Discussion: While the RLPTO approach maximizes L. An Integrated Federated Machine Learning and Blockchain rewards in a variety of task contexts by skillfully balancing Framework With Optimal Miner Selection for Reliable DDOS task urgency and resource allocation, its effectiveness can Attack Detection differ depending on the particulars of datasets and compute D. Saveetha, G. Maragatham, V. Ponnusamy and N. nodes. Subsequent research endeavors ought to delve deeper Zdravković into augmenting the scheduling efficacy of RLPTO and tackle 1) Introduction: Despite providing security and trans- plausible constraints associated with dynamic system modifi- parency, distributed denial of service (DDoS) assaults and cations. other forms of cyberattacks can still be launched against 5) Conclusion: This study shows how RLPTO might en- blockchain technology. Since traditional approaches are unable hance model training results by providing a novel method for to handle the growing volume and complexity of blockchain resource and task scheduling in distributed learning clusters. data, effective detection mechanisms become increasingly Subsequent research endeavors will concentrate on enhancing important. This study highlights how crucial it is to use the technique to guarantee even higher effectiveness in man- blockchain technology and federated machine learning to aging changing computational requirements. improve security overall and DDoS attack detection.  2) Methodology: The suggested method makes use of a significance of comparing models to confirm efficacy, and the blockchain network’s federated machine learning framework fact that most research evaluate performance using indicators to identify DDoS attacks by using the dispersed nodes (miners) such as true positive rate highlights the potential of machine to build local models using attack data. To guarantee optimal learning to improve cloud security. miner involvement and safeguard the integrity of the machine 4) Discussion: Although the review offers insightful infor- learning model—which is kept on the blockchain to avoid tam- mation about the use of machine learning in cloud security, it pering—a dynamic reputation-based miner selection method is also draws attention to the paucity of new datasets and the created. sparse application of deep learning techniques. The results 3) Key Findings: Using the Random Forest algorithm, point to the necessity of more empirical research and updated the blockchain-integrated federated learning approach out- datasets to investigate cutting-edge machine learning tech- performed other detection systems, achieving an astounding niques, especially with regard to improving model accuracy 99.1% accuracy in DDoS assault detection. This great degree and resolving imbalanced datasets. of precision demonstrates how well the suggested structure 5) Conclusion: This comprehensive overview of the lit- works to strengthen blockchain network security and depend- erature emphasizes how important machine learning is to ability against DDoS attacks. solving cloud security problems, especially those involving 4) Discussion: Although the suggested framework shows DDoS attacks and data privacy concerns. In order to further notable improvements in DDoS detection, it is mainly targeted develop cloud security solutions, it emphasizes the necessity at one particular kind of attack. To ensure complete security for updated methodology and diversified datasets, and it urges in blockchain environments, further research should examine for continued research in this area. mitigation measures to supplement detection efforts and ad- dress the framework’s applicability to other attack vectors. N. Distributed Denial of Service Attack Detection for the 5) Conclusion: By using a federated machine learning Internet of Things Using Hybrid Deep Learning Model methodology, this study offers a solid method for identi- A. Ahmim, F. Maazouzi, M. Ahmim, S. Namane and I. B. fying DDoS attacks inside blockchain networks, improving Dhaou the security and dependability of the blockchain and the machine learning models used. Future security protections 1) Introduction: Because real traffic patterns are so com- against a wider range of threats may be made possible by plicated, traditional machine learning techniques frequently this framework’s ongoing development and testing. fail to detect DDoS attacks. Designed for the Internet of Things, this study presents a revolutionary deep learning-based M. Machine Learning for Cloud Security: A Systematic Re- intrusion detection system with the goal of improving security view against various DDoS attacks. A. B. Nassif, M. A. Talib, Q. Nasir, H. Albadani and F. M. 2) Methodology: Convolutional neural networks Dakalbab (CNNs), long short-term memory networks (LSTMs), 1) Introduction: Significant security difficulties are brought deep autoencoders, and deep neural networks (DNNs) are just about by the cloud’s increasing adoption; as a result, robust a few of the neural network architectures that are integrated detection and prevention techniques are required to guard into the suggested model’s hybrid deep learning architecture. against a variety of threats, such as DDoS and privacy In order to enhance detection skills, this two-level model breaches. Techniques for machine learning (ML) have first processes input data using parallel sub-networks, which shown promise in addressing these security flaws. The ob- are then integrated in the second level. The CIC-DDoS2019 jective of this systematic literature review is to emphasize dataset, which covers a range of DDoS attack types, was used the efficacy of machine learning (ML) approaches in cloud to assess the model’s performance. security by consolidating knowledge and addressing new risks. 3) Key Findings: The suggested intrusion detection system 2) Methodology: After conducting a thorough examination produced a very low false alarm rate of 0.04%, an average de- of the literature, 63 pertinent papers that were published tection rate of 71.42%, and a worldwide accuracy of 80.75%. between 2004 and 2019 were examined. Three main research These findings show the model’s efficacy in differentiating topics were the focus of the review: what kinds of cloud between different DDoS attack subcategories in actual traffic security threats are addressed, what machine learning tech- scenarios, outperforming other well-known machine learning niques are applied, and what are the performance outcomes of and deep learning models. these techniques? This thorough approach made it possible to 4) Discussion: Although the results show notable improve- determine the most common security vulnerabilities and assess ments in DDoS detection in IoT systems, additional testing different machine learning techniques in cloud environments. on more recent and diverse datasets is necessary, as the model 3) Key Findings: Eleven different cloud security domains is heavily dependent on the CIC-DDoS2019 dataset. Future were identified by the review, with DDoS and data privacy improvements like feature selection and hyperparameter op- receiving the greatest attention. Thirty distinct machine learn- timization may also boost the model’s performance, even if ing approaches were examined, with Support Vector Machines the architecture as it is now has encouraging accuracy and (SVM) being the most widely used. The results highlight the detection rates.  5) Conclusion: The hybrid deep learning model presented learning capabilities. Due to the speed and scalability issues in this paper successfully tackles the problems associated with with traditional single-machine methodologies, distributed ma- DDoS attack detection in Internet of Things environments. chine learning frameworks that make use of high-performance The suggested approach shows notable increases in detection computing clusters have been developed. Nonetheless, capabilities by utilizing the advantages of many deep learning current techniques for aggregating models, including parame- techniques, laying the groundwork for future studies into ter averaging, sometimes jeopardize the accuracy of training, improving intrusion detection in progressively complex IoT particularly in situations involving nonconvex optimization. contexts. 2) Methodology: In order to maximize model parameter aggregation, this work presents a novel loss function weight O. Dynamic Replication Policy on HDFS Based on Machine reorder stochastic gradient descent method (LR-SGD) that Learning Clustering establishes node weights according to their respective loss M. A. Ahmed, M. H. Khafagy, M. E. Shaheen and M. R. Kaseb function values. LR-SGD seeks to increase training accuracy 1) Introduction: Efficient storage solutions are necessary for Stale Synchronous Parallel (SSP) and Bulk Synchronous due to the recent rapid development of data, especially in the Parallel (BSP) models in distributed situations by improving context of big data research and distributed file systems (DFS) the parameter averaging method. such as Hadoop Distributed File System (HDFS). HDFS’s 3) Key Findings: According to experimental findings, LR- default replication process offers great data availability and re- SGD can improve training accuracy for the BSP model by liability, but it also results in significant resource overhead and up to 0.57% and for the SSP model by about 6.30%. These inefficiencies. As a result, replication strategies that optimize enhancements are noteworthy because they show how well storage and performance are required. the technique works to optimize distributed learning for smart 2) Methodology: Using unsupervised learning approaches, sensing applications, which could result in improved real-time this study offers a Dynamic Replication Policy using Machine data processing speed. Learning Clustering (DRPMLC) that divides HDFS files into 4) Discussion: The results indicate that LR-SGD can be three clusters: Hot, Warm, and Cold, based on their relative a useful tool for addressing the drawbacks of conventional relevance. Then, distinct replication policies are implemented parameter averaging techniques, especially when dealing with for every cluster, enabling more effective storage utilization nonconvex optimization problems. To investigate its adaptabil- while preserving data availability. ity across many distributed learning environments and smart 3) Key Findings: In comparison to the default HDFS sensing applications, more research is necessary. Nevertheless, replication, the DRPMLC system reduced replication storage the dependence on loss function values may pose additional space by almost 50% and improved read and write operation complications. times by 28.2% and 29%, respectively. These findings show 5) Conclusion: The suggested LR-SGD approach improves how clustering can effectively optimize resource use in HDFS training efficiency and accuracy and is a major improvement environments—a critical aspect of large data management. in distributed machine learning for smart sensing devices. 4) Discussion: The results imply that machine learning- Subsequent investigations will endeavor to incorporate this driven clustering might greatly improve HDFS data replication methodology alongside other model aggregation strategies to performance, which may have wider implications for other enhance efficacy in a variety of applications, so clearing the distributed systems. However, as the current implementation is path for enhanced real-time data processing in intelligent dependent on certain clustering algorithms, future studies may technologies. need to examine how scalable and flexible these techniques are under other workload scenarios and data types. Q. The Research on Distributed Fusion Estimation Based on 5) Conclusion: Introducing the DRPMLC system is a big Machine Learning step toward using machine learning for dynamic clustering Z. Peng, Y. Li and G. Hao and optimizing data replication in HDFS. This strategy lowers 1) Introduction: Even if they are the best, traditional cen- storage costs while simultaneously speeding up data access, tralized fusion techniques have significant computational opening the door for later improvements in replication and and communication overhead, which makes distributed fusion clustering tactics to further improve distributed file system techniques more desirable for large-scale sensor networks. efficiency. This study proposes machine learning-based fusion approaches that use local estimations to increase accuracy without relying P. Model Aggregation Method for Data Parallelism in Dis- on complex covariance matrices or real states, thereby address- tributed Real-Time Machine Learning of Smart Sensing Equip- ing the shortcomings of current distributed fusion algorithms. ment 2) Methodology: The paper presents two artificial neural Y. Fan, Z. Wei, J. Zhang, N. Zhao, Y. Ren, J. Wan, L. Zhou, Z. network (ANN) based multi-sensor distributed fusion frame- Shen, J. Wang, and J. Zhang. works: one that uses an Elman network and the other that is 1) Introduction: The proliferation of intelligent sensing based on a Back Propagation (BP) network. While the Elman devices has produced enormous volumes of data, making network uses centralized fusion estimates as training sets to effective training techniques necessary to enhance machine achieve higher accuracy in situations when genuine states or cross-covariance matrices are not accessible, the BP network model; no statistically significant variations in area under the techniques train on traditional distributed fusion estimations curve (AUC) values were found (p ¿ 0.05). These results with local inputs. suggest that C-DistriM can produce results that are on par 3) Key Findings: According to simulation results, the sug- with conventional approaches while improving data security gested Elman net-based fusion algorithm outperforms con- and trust across involved institutions, which will encourage a ventional distributed fusion algorithms in terms of accuracy, wider use of AI in multicentric healthcare studies. especially in situations where genuine state information is 4) Discussion: By combining distributed learning with unavailable. This improvement demonstrates how machine blockchain technology, important issues of trust and trans- learning approaches may be used to optimize multi-sensor parency are addressed, enabling researchers to keep an eye fusion procedures, which will ultimately result in more de- on the integrity of model training and data provenance. How- pendable system performance in practical applications. ever, in order to enable broad adoption in various healthcare 4) Discussion: The results show that machine learning- contexts, there are certain restrictions that must be addressed, based fusion techniques can successfully overcome the draw- such as the requirement for a strong blockchain infrastructure backs of traditional algorithms, especially with regard to and certain scalability issues. their dependence on particular statistical criteria. However, 5) Conclusion: The C-DistriM method shows that it is some BP-based approaches may be difficult to execute in possible to combine distributed learning and blockchain tech- practice since they require real states and measurement noise nology to attain performance levels comparable to centralized covariance matrices. To increase the range of applications for models while also improving data handling transparency and the suggested methodologies, more investigation into these trust. With the help of more informed clinical decision-making, constraints is required. this novel approach may foster international cooperation in 5) Conclusion: The multi-sensor distributed fusion estima- AI-driven healthcare, ultimately leading to better patient out- tion algorithms that have been suggested make use of machine comes. learning techniques to enhance the precision of information fusion in sensor networks, especially in difficult situations. In S. A Machine Learning Auxiliary Approach for the Distributed order to further improve performance across a variety of sensor Dense RFID Readers Arrangement Algorithm applications, future research will concentrate on improving P. Yan, S. Choudhury and R. Wei these algorithms and investigating their integration with other 1) Introduction: RFID technology is essential for busi- machine learning approaches. nesses such as supply chain management; nevertheless, when numerous readers are operating in close proximity, tags col- R. Blockchain for Privacy Preserving and Trustworthy Dis- lide, posing a substantial obstacle to dense RFID systems. tributed Machine Learning in Multicentric Medical Imaging Even if they work well, the current centralized anti-collision (C-DistriM) algorithms add to the computational burden and complexity. Fadila Zerka, Visara Urovi, Akshayaa Vaidyanathan, Samir This emphasizes the need for more effective distributed meth- Barakat, Ralph T. H. Leijenaar, Sean Walsh, Hanif Gabrani- ods that can use machine learning to improve efficiency and Juma, Benjamin Miraglio, Henry C. Woodruff, Michel Dumon- lessen dependency on external data. tier, Philippe Lambin 2) Methodology: This work presents a machine learning 1) Introduction: The reliability of the models’ predictions model that improves a distributed anti-collision method by and the caliber of the data used to train them determine compensating for the lack of global knowledge. The algorithm how useful artificial intelligence (AI) will be in the health- is based on the centralized MWISBAII approach. Activation care industry. Alternative strategies, such as distributed scores for RFID readers are predicted using a multi-layer neu- learning, are becoming more and more necessary as legal and ral network that was trained with data from early simulations ethical issues impede traditional centralized approaches to data of the MWISBAII algorithm. This enables better informed sharing. However, issues with data quality and transparency decision making to occur during tag interrogation. in distributed frameworks force researchers to look at novel 3) Key Findings: Results from experiments show that un- approaches, such incorporating blockchain technology to im- der different experimental situations, the machine learning- prove traceability and trust. enhanced distributed method performs almost as well as the 2) Methodology: This paper presents a unique method centralized MWISBAII algorithm. This result emphasizes how called Chained Distributed Machine Learning (C-DistriM), well machine learning can be integrated into distributed sys- which combines a blockchain infrastructure with sequential tems to enable effective collision avoidance in crowded RFID distributed learning. The technique ensures that all interac- environments. tions are documented on the blockchain for transparency and 4) Discussion: Machine learning greatly enhances RFID permits various health centers to cooperatively train machine reader decision-making; yet, the method relies on a uniform learning models without transferring sensitive data using the reader distribution, which might not accurately represent real- NSCLC-Radiomics dataset. world situations. Future research should investigate unsuper- 3) Key Findings: In six distinct scenarios, the C-DistriM vised learning methods to more effectively classify readers framework’s performance was compared to a centralized according to their operational traits; however, issues like choosing the right cluster sizes and controlling communication works frequently ignore data integrity, leaving them open overhead need to be resolved. to attacks that could jeopardize training data. Maintaining 5) Conclusion: The study offers a viable path for enhancing data integrity is crucial since tampered data might result in anti-collision algorithms in dense RFID systems by using a inaccurate model training and untrustworthy AI applications. dispersed strategy that makes use of machine learning. This In order to close this important gap, this work suggests a work paves the path for more scalable and efficient RFID verification scheme that protects the training data integrity in applications in many industrial settings by reducing the per- DML systems. formance gap between centralized and distributed techniques. 2) Methodology: Provable Data Possession (PDP) sampling T. Machine Learning Approaches for Combating Distributed auditing algorithm is used in the proposed DML-DIV scheme Denial of Service Attacks in Modern Networking Environments to verify data integrity and detect tampering and forgery A. Aljuhani attempts. The discrete logarithm problem is used as a blinding factor to safeguard anonymity during third-party audits, while 1) Introduction: Attacks known as distributed denial of identity-based cryptography is employed to streamline key service (DDoS) present a serious risk to service providers management and lessen key escrow concerns. because they flood systems with malicious requests, which halt 3) Key Findings: The DML-DIV method lowers certificate operations and cause large financial losses. Effective detection management expenses, protects privacy, and efficiently assures and mitigation solutions are urgently needed as attackers use the integrity of training data. The promise of DML-DIV to more complex approaches, such as DDoS as a service improve the security of distributed machine learning frame- (DDoSaaS). Machine learning (ML) techniques can play a works is highlighted by its performance above current integrity major role in this process. verification techniques, as demonstrated by both theoretical 2) Methodology: The DDoS detection techniques that use analysis and simulation results. single and hybrid machine learning approaches are reviewed in 4) Discussion: The security of training data in DML sys- this paper based on recent research. It looks at different DDoS tems can be effectively addressed by using DML-DIV, which defensive solutions in contemporary networking settings, such enhances the dependability of AI applications. Scalability as cloud computing, network functions virtualization (NFV), and real-time performance issues with the approach could and software-defined networks (SDN), and it also covers the arise, though, particularly in settings with high data volumes. particular difficulties in protecting Internet of Things (IoT) Subsequent efforts may concentrate on refining the plan for devices against DDoS attacks. wider implementation. 3) Key Findings: The investigation shows that the capacity 5) Conclusion: This study addresses important security of security systems to identify and counter DDoS attacks in a issues by presenting a thorough data integrity verification variety of settings is much improved by machine learning ap- scheme designed for distributed machine learning environ- proaches. Through the classification of different DDoS attack ments. The DML-DIV scheme is a crucial development for types and the ML strategies that go along with them, the study the safe implementation of machine learning technologies in highlights how cybersecurity is changing and how crucial it is a range of applications since it improves data integrity and to modify these approaches for use with contemporary network privacy. infrastructures. 4) Discussion: Although the paper underlines persistent V. Open-Set Recognition in Unknown DDoS Attacks Detection issues such attacker adaptation and the need for continued With Reciprocal Points Learning detection algorithm enhancement, ML-based protection sys- C. -S. Shieh, F. -A. Ho, M. -F. Horng, T. -T. Nguyen and P. tems demonstrate potential. Furthermore, in order to maintain Chakrabarti effective security measures, it is necessary to handle the 1) Introduction: By overloading network resources, dis- opportunities and complications presented by the fast rise of tributed denial-of-service (DDoS) attacks pose a serious dan- virtualized systems. 5) Conclusion: The study comes to the conclusion that ger to cybersecurity. They can cause major disruptions to DDoS attacks are still a serious concern and that creative online services. Because they rely on known attack signa- mitigation measures utilizing ML and deep learning ap- tures, traditional intrusion detection systems (IDS) find it proaches are required. It highlights the significance of cre- difficult to identify unknown DDoS assault forms. This under- ating strong defense systems that are appropriate for today’s scores the critical need for more flexible detection techniques. networking environments and identifies important research In order to overcome these obstacles, this study suggests a avenues to strengthen service providers’ resilience against novel approach that improves detection capacities for both these widespread attacks. known and unidentified DDoS attacks. 2) Methodology: Using Open-Set Recognition (OSR) ap- U. Distributed Machine Learning Oriented Data Integrity proaches, the proposed CNN-RPL model integrates Reciprocal Verification Scheme in Cloud Computing Environment Points Learning (RPL) with Convolutional Neural Networks X. -P. Zhao and R. Jiang (CNN). This method improves the model’s generalization and 1) Introduction: Artificial intelligence (AI) relies heav- detection accuracy by taking attributes from DDoS data and ily on distributed machine learning (DML), yet current frame- using them to identify known attack types. At the same time, it adapts to novel attacks by limiting the distribution of existing 4) Discussion: Even while machine learning and its related categories. fields have a lot of potential for use in cybersecurity, there are 3) Key Findings: Impressive accuracy rates are attained by still obstacles to overcome, such as the requirement for high- the CNN-RPL model; in the CICIDS2017 dataset, it surpassed quality data and interpretable models. Additionally, enterprises 99.93% for known DDoS assaults, while in the CICDDoS2019 must manage these complexities to optimize advantages while dataset, it averaged 98.51% against unknown attacks. These minimizing risks because AI solutions like ChatGPT have the outcomes highlight the efficacy and efficiency of the model, potential to both strengthen and jeopardize cybersecurity. which has fewer parameters to improve operational flexibility 5) Conclusion: To sum up, in order to combat the al- without sacrificing defense capabilities. ways changing array of cyberthreats, cybersecurity strategies 4) Discussion: Even if the CNN-RPL model shows notable must incorporate ML, DL, RL, and AI capabilities. In an improvements in DDoS detection, there are still issues to be increasingly digital world, it is imperative that research and resolved, especially when it comes to making adjustments development in these areas continue in order to fortify defenses to the always changing landscape of attack methods and and efficiently preserve sensitive data. guaranteeing a workable implementation in a variety of net- work scenarios. To further improve detection accuracy, future developments might concentrate on diversifying datasets and X. Machine Learning for Security and the Internet of Things: improving Open-Set identification techniques. The Good, the Bad, and the Ugly 5) Conclusion: This work greatly increases the adaptability F. Liang, W. G. Hatcher, W. Liao, W. Gao and W. Yu of intrusion detection systems by introducing a reliable and effective model for identifying unknown DDoS attacks. The 1) Introduction: Cyber-physical systems (CPS) have been CNN-RPL model is a useful tool for improving cybersecurity revolutionized by the quick development of the Internet of defenses since it not only performs well in thwarting existing Things (IoT) and machine learning, yet these develop- attacks but also demonstrates the ability to handle novel and ments have also brought forth serious risks. The rising tar- developing threats. geting of these vital systems by malevolent actors who take advantage of unanticipated flaws highlights the critical need W. A Comprehensive Survey: Evaluating the Efficiency of to comprehend machine learning’s dual role in cybersecu- Artificial Intelligence and Machine Learning Techniques on rity—that of a safeguard and a possible weapon. Cyber Security Solutions 2) Methodology: This research uses a framework to exam- Ozkan-Ozay, Merve, Erdal Akin, Ömer Aslan, Selahattin Ko- ine the ways in which machine learning is used in cybersecu- sunalp, Teodor Iliev, Ivaylo Stoyanov, and Ivan Beloev. rity and CPS, classifying its applications into three areas: the ”Good,” which highlights applications that are advantageous; 1) Introduction: Advanced security measures are required the ”Bad,” which highlights attacks that pose a threat to due to the considerable hazards that the surge in cyberat- security; and the ”Ugly,” which looks at how machine learning tacks poses to individuals, organizations, and governments. In is used to enable cyberattacks. This analysis is informed by order to improve detection and response capabilities, novel an extensive survey of current literature and methodologies. approaches like machine learning (ML), deep learning (DL), and reinforcement learning (RL) are required. Conventional 3) Key Findings: According to the research, machine learn- cybersecurity methods frequently fall short in identifying ing improves security and decision-making in CPS, but it sophisticated assaults. Leveraging these technologies is also has serious flaws that might be exploited. The study essential for protecting digital assets as assaults get more emphasizes how crucial it is to secure machine learning sophisticated and data volumes rise. systems to stop bad actors from turning them into weapons and 2) Methodology: This research assesses the use of machine how strong defensive tactics are against adversarial attacks. learning (ML), deep learning (DL), and reinforcement learning 4) Discussion: The findings of this study underline how (RL) techniques in cybersecurity by looking at their designs urgently machine learning applications need to improve their and efficacy in a range of domains, including vulnerability defensive mechanisms in order to fend off both established assessment, malware detection, and intrusion detection. The and new dangers. The report does, however, admit many study also examines cutting-edge research, stressing important shortcomings, most notably the absence of current protections discoveries and the difficulties in putting these AI-driven against machine learning-based attacks and the continuous strategies into practice. need for additional research to address these vulnerabilities. 3) Key Findings: According to the research, ML, DL, and 5) Conclusion: Finally, this study highlights how machine RL greatly enhance cyber threat detection and mitigation, learning in cybersecurity and CPS has two sides, highlighting allowing for quicker and more precise attack reaction times. how it can both improve and jeopardize system integrity. Although there are many advantages to these methods, the Developing successful tactics to protect critical infrastructures study also highlights drawbacks, such as the possibility of from emerging cyber threats requires a thorough understanding false positives and poor data quality, which may limit their of these dynamics, which calls for continuous study and usefulness in practical settings. innovation in defensive technologies. Y. Power Allocation Schemes Based on Machine Learning for The comparison of machine learning techniques with con- Distributed Antenna Systems ventional optimization methods demonstrates the usefulness Y. Liu, C. He, X. Li, C. Zhang and C. Tian of using models like k-nearest neighbors (k-NN) in DAS, 1) Introduction: Efficient resource allocation algorithms which drastically lower computational complexity without are imperative in cellular networks due to the exponential compromising performance. But it’s imperative to overcome growth of data traffic, especially in distributed antenna sys- the shortcomings of the approaches used today, especially the tems (DAS). Real-time applications cannot effectively use reliance on reliable historical data. traditional power allocation optimization algorithms due to their high computational complexity. This research investi- R EFERENCES gates how machine learning—more specifically, the k-NN algorithm—can be integrated to provide a more workable and G. Qian, Z. Li, C. He, X. Li and X. Ding, ”Power Allocation Schemes Based on Deep Learning for Distributed Antenna Systems,” effective solution for DAS power allocation. in IEEE Access, vol. 8, pp. 31245-31253, 2020, doi: 10.1109/AC- 2) Methodology: A new system model utilizing the k- CESS.2020.2973253. M. N. H. Nguyen et al., ”Self-Organizing Democratized Learning: To- NN algorithm and historical data from conventional optimiza- ward Large-Scale Distributed Learning Systems,” in IEEE Transactions tion techniques—more especially, the sub-gradient iterative on Neural Networks and Learning Systems, vol. 34, no. 12, pp. 10698- method—is presented in the study. The method evaluates the 10710, Dec. 2023, doi: 10.1109/TNNLS.2022.3170872. performance of the k-NN algorithm against well-established H. Afifi et al., ”Machine Learning With Computer Networks: Tech- niques, Datasets, and Models,” in IEEE Access, vol. 12, pp. 54673- traditional approaches by comparing power allocation schemes 54720, 2024, doi: 10.1109/ACCESS.2024.3384460. for maximizing spectral efficiency (SE) and energy efficiency Y. Chen, Y. Mao, H. Liang, S. Yu, Y. Wei and S. Leng, ”Data Poison (EE). This ensures decreased computational complexity with- Detection Schemes for Distributed Machine Learning,” in IEEE Access, vol. 8, pp. 7442-7454, 2020, doi: 10.1109/ACCESS.2019.2962525. out sacrificing accuracy. M. Aristodemou, X. Liu, S. Lambotharan and B. AsSadhan, ”Bayesian 3) Key Findings: According to simulation studies, the Optimization-Driven Adversarial Poisoning Attacks Against Distributed power allocation schemes that come from the k-NN algo- Learning,” in IEEE Access, vol. 11, pp. 86214-86226, 2023, doi: 10.1109/ACCESS.2023.3304541. rithm closely resemble those that originate from conventional E. Zeydan, S. S. Arslan and M. Liyanage, ”Managing Distributed techniques, producing SE and EE results that are comparable. Machine Learning Lifecycle for Healthcare Data in the Cloud,” in This proves that DAS’s resource allocation procedure can be IEEE Access, vol. 12, pp. 115750-115774, 2024, doi: 10.1109/AC- CESS.2024.3443520. efficiently streamlined with machine learning, improving its A. Aminifar, M. Shokri, F. Rabbi, V. K. I. Pun and Y. Lamo, ”Extremely suitability for dynamic, real-world applications. Randomized Trees With Privacy Preservation for Distributed Structured 4) Discussion: The results point to important ramifications Health Data,” in IEEE Access, vol. 10, pp. 6010-6027, 2022, doi: 10.1109/ACCESS.2022.3141709. for next wireless communication systems and demonstrate P. Shen, X. Du and C. Li, ”Distributed Semi-Supervised Metric Learn- how machine learning can improve power allocation efficiency. ing,” in IEEE Access, vol. 4, pp. 8558-8571, 2016, doi: 10.1109/AC- There are still issues, though, mainly with the dependence on CESS.2016.2632158. A. L’Heureux, K. Grolinger, H. F. Elyamany and M. A. M. Capretz, the quality of past data and the want for more investigation ”Machine Learning With Big Data: Challenges and Approaches,” into different machine learning methods that can enhance in IEEE Access, vol. 5, pp. 7776-7797, 2017, doi: 10.1109/AC- performance even more. CESS.2017.2696365. X. Lu, C. Liu, S. Zhu, Y. Mao, P. Lio and P. Hui, ”RLPTO: A 5) Conclusion: This research concludes by demonstrating Reinforcement Learning-Based Performance-Time Optimized Task and the promising role of machine learning in enhancing DAS Resource Scheduling Mechanism for Distributed Machine Learning,” in power distribution tactics, obtaining comparable outcomes IEEE Transactions on Parallel and Distributed Systems, vol. 34, no. 12, pp. 3266-3279, Dec. 2023, doi: 10.1109/TPDS.2023.3317388. to conventional approaches with notably reduced computing N. S. Musa, N. M. Mirza, S. H. Rafique, A. M. Abdallah and T. complexity. The work establishes the foundation for further Murugan, ”Machine Learning and Deep Learning Techniques for Dis- investigations into various machine learning approaches with tributed Denial of Service Anomaly Detection in Software Defined the goal of improving effectiveness and usefulness in wireless Networks—Current Research Solutions,” in IEEE Access, vol. 12, pp. 17982-18011, 2024, doi: 10.1109/ACCESS.2024.3360868. communication networks. D. Saveetha, G. Maragatham, V. Ponnusamy and N. Zdravković, ”An Integrated Federated Machine Learning and Blockchain Framework III. C ONCLUSION With Optimal Miner Selection for Reliable DDOS Attack Detection,” in IEEE Access, vol. 12, pp. 127903-127915, 2024, doi: 10.1109/AC- We have looked at the relationship between machine learn- CESS.2024.3413076. A. B. Nassif, M. A. Talib, Q. Nasir, H. Albadani and F. M. Dakalbab, ing and distributed systems in this literature review, with a ”Machine Learning for Cloud Security: A Systematic Review,” in particular emphasis on how it applies to distributed antenna IEEE Access, vol. 9, pp. 20717-20735, 2021, doi: 10.1109/AC- systems (DAS) and cybersecurity. The analysis shows that al- CESS.2021.3054129. though machine learning greatly improves system performance A. Ahmim, F. Maazouzi, M. Ahmim, S. Namane and I. B. Dhaou, ”Distributed Denial of Service Attack Detection for the Internet of and optimizes resource allocation, it also brings new risks and Things Using Hybrid Deep Learning Model,” in IEEE Access, vol. 11, difficulties. Because machine learning may be used for both pp. 119862-119875, 2023, doi: 10.1109/ACCESS.2023.3327620. good and ill, it is dual in nature, which emphasizes the need M. A. Ahmed, M. H. Khafagy, M. E. Shaheen and M. R. Kaseb, ”Dynamic Replication Policy on HDFS Based on Machine Learning for strong defenses against adversarial assaults in cybersecurity Clustering,” in IEEE Access, vol. 11, pp. 18551-18559, 2023, doi: scenarios. 10.1109/ACCESS.2023.3247190.  Y. Fan et al., ”Model Aggregation Method for Data Parallelism in Distributed Real-Time Machine Learning of Smart Sensing Equipment,” in IEEE Access, vol. 7, pp. 172065-172073, 2019, doi: 10.1109/AC- CESS.2019.2955547. Z. Peng, Y. Li and G. Hao, ”The Research on Distributed Fusion Estimation Based on Machine Learning,” in IEEE Access, vol. 8, pp. 38174-38184, 2020, doi: 10.1109/ACCESS.2020.2974039. F. Zerka et al., ”Blockchain for Privacy Preserving and Trustworthy Distributed Machine Learning in Multicentric Medical Imaging (C- DistriM),” in IEEE Access, vol. 8, pp. 183939-183951, 2020, doi: 10.1109/ACCESS.2020.3029445. P. Yan, S. Choudhury and R. Wei, ”A Machine Learning Auxil- iary Approach for the Distributed Dense RFID Readers Arrangement Algorithm,” in IEEE Access, vol. 8, pp. 42270-42284, 2020, doi: 10.1109/ACCESS.2020.2977683. A. Aljuhani, ”Machine Learning Approaches for Combating Distributed Denial of Service Attacks in Modern Networking Environments,” in IEEE Access, vol. 9, pp. 42236-42264, 2021, doi: 10.1109/AC- CESS.2021.3062909. X. -P. Zhao and R. Jiang, ”Distributed Machine Learning Oriented Data Integrity Verification Scheme in Cloud Computing Environment,” in IEEE Access, vol. 8, pp. 26372-26384, 2020, doi: 10.1109/AC- CESS.2020.2971519. C. -S. Shieh, F. -A. Ho, M. -F. Horng, T. -T. Nguyen and P. Chakrabarti, ”Open-Set Recognition in Unknown DDoS Attacks Detection With Reciprocal Points Learning,” in IEEE Access, vol. 12, pp. 56461-56476, 2024, doi: 10.1109/ACCESS.2024.3388149. M. Ozkan-Okay et al., ”A Comprehensive Survey: Evaluating the Efficiency of Artificial Intelligence and Machine Learning Techniques on Cyber Security Solutions,” in IEEE Access, vol. 12, pp. 12229-12256, 2024, doi: 10.1109/ACCESS.2024.3355547. F. Liang, W. G. Hatcher, W. Liao, W. Gao and W. Yu, ”Machine Learning for Security and the Internet of Things: The Good, the Bad, and the Ugly,” in IEEE Access, vol. 7, pp. 158126-158147, 2019, doi: 10.1109/ACCESS.2019.2948912. Y. Liu, C. He, X. Li, C. Zhang and C. Tian, ”Power Allocation Schemes Based on Machine Learning for Distributed Antenna Systems,” in IEEE Access, vol. 7, pp. 20577-20584, 2019, doi: 10.1109/AC- CESS.2019.2896134.