Cloud Computing: Security, Resource Optimization, and Algorithm Evaluation

Research Questions

This study aims to present the results of a systematic review of the scientific literature on cloud computing, with emphasis on security, resource optimization, algorithm evaluation, and service performance. To construct the state of the art, the following research questions were formulated:

RQ1: What metrics and approaches make it possible to ensure optimal performance levels in cloud computing services?

RQ2: What factors and methodologies are used to evaluate algorithm performance in cloud computing environments?

RQ3: What tools and strategies are used to ensure the integrity, privacy, and transparency of data in the cloud?

RQ4: What metaheuristic algorithms have been proposed to optimize task scheduling and allocation in cloud environments?

RQ5: What virtual machine consolidation techniques are used to improve resource utilization in cloud data centers?

RQ6: What are the main security challenges in cloud computing infrastructure?

These questions were formulated based on a preliminary analysis of the literature and structured under a systematic approach that made it possible to delimit the scope of the study and guide the process of searching, selecting, and analyzing the information.

Search

The proposed search strategy consisted of exploring general terms in specialized databases. In the selected databases, search terms combined with synonyms were applied to cover a greater number of documents to be evaluated. Table 2 presents the terms with their respective synonyms and additional filters.

The search process was conducted in the databases mentioned in Table 3, using fields such as title, keywords, abstract, and full document. Table 4 presents the resulting search string for the main term, together with its synonyms and corresponding filters.

It should be noted that the search was limited to 2021 onward, since this period concentrates the most recent and relevant academic and scientific contributions on cloud computing, particularly regarding security, resource optimization, and algorithm evaluation in distributed environments, areas that have experienced significant growth in scientific production validated by the international research community.

Table 2. Term, synonyms, and filters for composing the search string

Source: own elaboration.

Table 3. Main term and database.

Source: own elaboration.

Table 4. Search string.

Source: own elaboration.

Selection

All studies included in this review were analyzed; during their evaluation, the following items were considered: title, keywords, abstract, introduction, background, state of the art, methodology, results, and conclusions. Likewise, inclusion and exclusion criteria were defined to guide the review.

The inclusion criterion defined is: a scientific document related to studies on security, resource optimization, algorithm evaluation, and service performance in cloud computing environments, or their search synonyms. The exclusion criteria defined are: studies that do not have their corresponding bibliographic citation, documents that do not contain the search terms or synonyms, works that do not directly address cloud computing environments, and documents that are not available for full download.

Table 5. Dates of consultation and file download.

Source: own elaboration.

Quality assessment

In the quality assessment process for the selected documents, 7 criteria were considered: provenance of the sources, relevance of the content, impact of the study, objective of the research, context of the study, objectivity of the methodological design, and scientific rigor in data analysis. These 7 criteria correspond to 3 fundamental elements of quality management: planning, organization, and control.

The selection of primary sources was carried out in 5 stages, corresponding to the 5 specialized databases consulted, each with 3 phases: phase 1, elimination of duplicate articles; phase 2, elimination of non-downloadable articles; and phase 3, application of the inclusion and exclusion criteria. Because this systematic review used 5 databases with their respective synonyms and filters —in each of which the 3 aforementioned phases were applied—, it was necessary to conduct the searches on different dates, as shown in Table 5.

As a result of the search process in the databases, a total of 120 studies were found; after applying the inclusion and exclusion criteria, a total of 68 relevant documents were obtained, which constitute the corpus of analysis for this systematic literature review.

Table 6. Quality assessment in search and selection processes.

Source: own elaboration.

Table 7. Studies included in the systematic review.

Source: own elaboration.

Factors and methodologies for evaluating the performance of algorithms in cloud computing environments

The performance of cloud algorithms is evaluated through multidimensional analysis that integrates makespan, throughput, resource utilization, energy expenditure, and load distribution balance [5]. The predominant factors span economic dimensions (cost, profit), operational dimensions (waiting time, failure rate, makespan), and environmental dimensions (carbon emissions, energy consumption) [33]. Single-parameter evaluation is ineffective; integrating completion time, energy consumption, resource utilization, and computational simulation cost simultaneously has been proposed [20]. For load balancing, temporal efficiency (total and average makespan), processing speed, and distribution effectiveness are measured [19]. Consolidation evaluation is based on energy consumption, SLA violations, VM migration count, and performance degradation [11].

Performance evaluation is frequently conducted through comparative analyses with state-of-the-art algorithms using QoS parameters, turnaround time, response time, and statistical tests to confirm the significance of the results [34]. For failure-prone environments, simulators such as iFogSim and paired t-tests are used to validate improvements against baselines [23]. The performance improvement rate (PIR) quantitatively compares new proposals with the state of the art, complemented by p-value analysis on NASA, HPC2N, and synthetic datasets [24].

In security, algorithms are evaluated with Silhouette and Davies-Bouldin indices, using bootstrap to derive 95% confidence intervals on precision and macro F1 [35]. CloudSim Plus simulates distributed infrastructures by measuring utilization, makespan, average latency, and SLA compliance with multiple runs with random seeds [18]. Predictive model validation uses Google Cluster traces with RMSE, F1, and AUC metrics to ensure reliability in fault detection [36]. Load prediction evaluation combines predictive accuracy (MAE, MSE) with deployment efficiency measured by complexity and inference time [37]. Experimental methodologies compare algorithms against baseline scenarios and optimal theoretical limits, focusing on energy savings, number of VM migrations, and SLA compliance [10].

A rigorous performance evaluation requires profiling tools such as Linux Perf or Intel VTune to identify microarchitectural bottlenecks, prioritizing queue latency and throughput in cloud environments [8]. Performance is comprehensively evaluated through algorithm reliability, monitoring of CPU, RAM, storage, and network traffic, and measurement of real-time response time [6]. Comprehensive system evaluation requires resource monitoring tools and benchmarks to standardize elasticity, latency, and performance tests [38]. Algorithms in container environments are evaluated with multi-objective analysis considering resource waste, load balancing, fragmentation, energy consumption, and convergence time with SLA compliance [13].

VM placement algorithms are measured by total energy consumption and resource utilization rate integrated into a fitness function (fitness score) [14]. Load balancing performance is quantified through the degree of imbalance (DOI) and Jain's fairness index in simulated infrastructures with heterogeneous capacities [21]. The comparison of TLBO, PSO, and GA evaluates cost-time balance, convergence speed, and scalability under increasing tasks [17].

Privacy-oriented evaluation uses a dual index: model accuracy and encryption cost, along with load latency and aggregation time [39]. Normative frameworks such as ISO/IEC 25010 analyze availability, reliability, and security, in addition to provider responsiveness [7]. Methodologies for rating services identify functional and non-functional KPIs, complemented by the CRITIC method to weight factors and Vikor to establish rankings [4]. For auto-scaling systems, metrics of allocation waiting time, service delay, and provisioning failures are prioritized [26]. Cloud performance analysis is mainly conducted through experiments, surveys, case studies, and literature reviews to ensure empirical validity [40].

Tools and strategies to ensure the integrity, privacy, and transparency of data in cloud environments

A comprehensive strategy in edge-cloud environments involves container hardening through digital signature verification and a SIEM system for real-time monitoring, ensuring that only authorized images process confidential data [41]. Federated learning guarantees privacy by keeping data local and sharing only model parameters through encrypted channels (VPN/TLS) and virtual private clouds (VPC), reducing the attack surface [42]. In serverless computing, integrity and privacy are ensured with encryption at rest and in transit, cryptographic signatures for event authentication, the principle of least privilege in IAM, and strict runtime isolation [29].

Transparency in the face of possible corruption requires combining encryption technologies with robust authentication mechanisms to verify the authorship of information [4]. Blockchain networks provide transparency through one decentralized and immutable ledger, complemented by homomorphic encryption that allows encrypted data to be processed without exposing its original content [43]. SDN architectures facilitate granular security policies with elliptic curve cryptography, hash functions, and digital signatures to detect tampering during transit [44]. Certificateless signcryption schemes unify digital signature and encryption, eliminating vulnerabilities of traditional certificates and key escrow risks in high-speed communications [45].

Blockchain technology with Merkle hash trees and smart contracts guarantees integrity and operational transparency; zero-knowledge proofs (ZKP) and secure multi-party computation (SMPC) verify authenticity without compromising confidentiality [46]. In serverless, sandboxing, memory erasure between executions, and encryption at rest and in transit, together with secure key management, enable compliance with GDPR and HIPAA [29]. Models for offloading data to the cloud require bidirectional mutual authentication combining SMS and image verification, together with proxy re-encryption and lightweight homomorphic encryption to efficiently process encrypted information [47]. Advanced security tools incorporate differential privacy and scaled attention layers in their AI models to detect threats without exposing users' identities [48]. Adoption in SMEs must be based on ISO 27005 for risk management and on CASB tools to control access and ensure compliance with privacy policies [7]. ISO 30300 and ISO 30301 standards define principles for secure recording and storage, facilitating the transition to encrypted digital archives with continuous availability [49].

The HFHE-Cloud model integrates federated learning and homomorphic encryption for multi-tenant environments, enabling collaborative training without exposing original data and ensuring traceability through on-chain auditing [39]. Protection in multi-tenant databases is achieved with automated CASE tools, logical isolation, RBAC, and partitioning policies to mitigate unauthorized access [50].

The Extended Zero Trust (EZT) methodology combines AI-based anomaly detection, continuous identity validation, and micro-segmentation to provide adaptive defense against insider threats [51]. DFL-RUL employs decentralized federated learning, transmitting only parameter updates, eliminating centralization of raw data, and complying with international data protection standards [52]. IoT Hub unifies data models and ensures consistency, relying on cloud services with advanced threat mitigation and regulatory compliance for transport and storage [53]. The Tree-based-Signature and CyberGuarder protocols secure virtual environments and are complemented by cryptography in Li-Fi systems to strengthen client privacy [40].

XAI tools with SHAP values enable interpretable perspectives on resource allocation, overcoming the black-box nature of traditional algorithms [54]. Rigorous access controls, identity management, and encryption techniques, together with validation of disaster recovery capabilities, ensure the continuous integrity of digital assets [55]. Redundancy and multiple backups guarantee integrity in distributed environments, while virtualization provides transparency to the end user by abstracting the physical location of processing [32]. Provable data possession techniques verify integrity in the cloud; they are critical in sectors with sensitive data, such as healthcare, and facilitate digital forensic audits [56]. Advanced encryption and automatic backups are the main strategies of the cloud operating framework for guaranteeing information security and privacy [11].

Metaheuristic algorithms for optimizing task scheduling and allocation in cloud environments

Metaheuristic methods are distinguished by their capacity to explore complex solution spaces; recognized approaches include GA, KHA, and CSA, as well as hybrid models such as PSO-SA and F-ACO to minimize makespan, energy consumption, and operating costs [57]. In addition to ACO, GA, and PSO, WOA, LOA, CS, and GWO have been proposed; these outperform traditional techniques in makespan and deadline violation rate [9]. In multi-objective programming, common techniques are ACO, GA, and PSO, while Cat Swarm Optimization has emerged to manage execution time and load in highly complex scenarios [5].

In edge and cloud environments, GA, FA, and PSO are employed as standard options, along with specialized variants such as OHNSGA and hybrid memetic algorithms (MADCP), to improve exploration and accelerate convergence [58]. For sustainable consolidation of VMs, IGA-POP, the Imperialist Competitive Algorithm (ICA), and Advanced Ant Colony Optimization (AACO) are reported, often integrated with machine learning techniques [10]. Sustainable management can be optimized using bio-inspired metaheuristics integrated into game theory frameworks such as GTGWO, PSO-GTA, and KASIA to maximize the use of clean energy [59].

The hybrid algorithm HHO-GWO integrates the global search of Harris Hawks Optimization with the local exploitation of Grey Wolf Optimizer for resource scheduling in containers, outperforming FA and WOA [13]. For scientific workflows, the DSSSO algorithm combines Dolphin Swarm Optimization (DSO) and Sparrow Search Algorithm (SSA), complemented by ACO and IBOA to optimize makespan and energy consumption [16]. GWO-EFO optimizes resource allocation in heterogeneous environments, while CSO and NSGA-III balance load and improve QoS [60]. The hybrid algorithm HCS-BA combines the Lévy flights of Cuckoo Search with the fine-tuning of the Bat Algorithm, while also reviewing Tabu search and FPO to reduce resource waste [14].

Traditional approaches employ the improved Multiverse Optimizer and adaptive PSO (AdPSO), which seek near-optimal solutions through dynamic adjustments to changes in load [25]. Simulated Annealing and genetic algorithms iteratively explore the solution space to avoid overprovisioning [61]. TLBO stands out for its fast convergence and low variance, whereas PSO is more effective in highly adaptable resource environments [17]. FireBat integrates the capabilities of Firefly with BAT, also categorizing PSO, ACO, ABC, and GWO as fundamental tools for reducing makespan [21].

The ICSO algorithm (Improved Chicken Swarm Optimization) documents effective metaheuristics, including algorithms inspired by whales, bats, cats, ant lions, and bee colonies for NP-hard problems [62]. HGHHC hybridizes HGSO and HHO by incorporating comprehensive opposition-based learning (COBL) to avoid local optima and improve makespan and resource usage [24]. GLOPS combines PSO for global exploration and Simulated Annealing for local refinement, achieving a 48.79% reduction in turnaround time compared to traditional approaches [22]. DALRS integrates GA for global refinement of assignments proposed by a reinforcement learning agent, managing more than 10,000 concurrent tasks while optimizing latency, cost, and energy consumption [18].

PSO offers lower costs and faster execution, whereas GA presents advantages in energy efficiency and reduction of operational costs under certain conditions in fog computing [63]. GIGOL integrates GA with GWO for multi-objective assignment problems, balancing response time, load balancing, and availability through population-based search and opposition-based learning [64]. In robust systems, MHHO+ACO balances load with fault awareness, improving the handling of critical tasks without excessive replication costs [23].

The research is divided between individual metaheuristics (bat, whale, cat swarm, ant lion) and hybridization strategies that combine ACO with PSO to maximize resource utilization [62]. PSO and ACO predominate in resource management; other approaches include cuckoo search, penguin optimization, firefly algorithms, and grey wolf optimization to reduce energy consumption and migrations [11]. MVO and CSO gain relevance due to their ability to handle the complexity of task scheduling; the integration of ACO and PSO enables dynamic load distribution adapted to real-time energy metrics [15]. For multi-objective optimization, SPEA, NPGA, and ABC are used, imitating biological behaviors to find optimal solutions in massive load environments [33]. The combination of GA with the Flower Pollination Algorithm (FPA) leverages the convergence of FPA and the crossover and mutation operators of GA to optimize completion time [20]. WOA and CSA overcome the limitations of traditional heuristics by avoiding local optima during task allocation [65].

Virtual machine consolidation techniques to optimize resource usage in cloud data centers

VM consolidation is a fundamental optimization method that groups multiple virtual instances onto a minimum number of physical servers; the most common techniques are Best-Fit Decreasing (BFD) and its modified variant (MBFD), which seek compact and efficient allocation by reducing waste and energy consumption [11]. The most effective techniques include Best Fit and E-MBFD policies integrated with regression models to predict CPU demand, enabling dynamic allocation with safety margins to absorb fluctuations without compromising SLAs [10]. In sustainable computing, game theory frameworks model VMs and hosts as strategic players; the GTGWO algorithm facilitates intelligent migration to data centers with renewable energy [59].

Quantum computing techniques and genetic algorithms solve multi-objective optimization problems, achieving utilization rates of up to 90% through load balancing that minimizes energy waste [5]. Push and Pull migration techniques based on linked lists allow VMs to adjust their load in real time, moving processes from overloaded nodes to those with idle capacity to guarantee optimal utilization and deadline compliance [9]. The hybrid HCS-BA algorithm balances global exploration and local exploitation to avoid resource waste from inactive servers and reduce the energy footprint of data centers [14]. Multiplexing and consolidation of workloads allow more tasks to be executed on the allocated capacity, optimizing physical infrastructure and maximizing benefits with fewer data centers [61].

Persuasive autoscaling (PASM) with deep reinforcement learning facilitates a more adaptable resource distribution by verifying shared instances and executing migrations based on the demand ratio to minimize waiting times and service failures [26]. Intelligent schedulers with GWO-EFO ensure optimal resource mapping, reducing execution times and guaranteeing maximum utilization of available physical and virtual capacity [60]. VM consolidation, combined with methods such as dynamic voltage and frequency scaling (DVFS), balances operational performance with power consumption [18]. Dynamic consolidation reorganizes workloads to optimize the utilization of available resources without degrading agreed service levels [22].

Load balancing and server consolidation are pillars for maximizing system efficiency by enabling an equitable distribution of workload and reducing underutilization through the grouping of virtual instances [21]. Algorithms such as ChicWhale facilitate VM migration, constituting a fundamental enabling component for consolidation processes that improve data management and efficiency in distributed resources [66]. The joint consolidation of VMs and containers enables more flexible and efficient management in green cloud environments, significantly reducing energy consumption [15].

Hybrid techniques that combine K-Means with reinforcement learning make it possible to accurately identify which VMs should be migrated from overloaded servers, guaranteeing optimal load balancing [67]. The skewness metric quantifies the inequality in the use of hardware resources, facilitating the consolidation of heterogeneous workloads to maximize overall utilization and support green computing [54]. Live migration concentrates workloads on fewer active nodes, making it possible to shut down idle servers without degrading the quality of service to the end customer [40].

Main security challenges in cloud infrastructure

The integration of edge-cloud infrastructures faces critical risks derived from its open architecture, including malicious data alteration and compute power hijacking, compromising availability and compliance with performance standards [41]. In serverless computing, shared infrastructure and event-based invocation amplify the challenges: gaps in tenant isolation lead to data leaks, and third-party libraries with permissive configurations facilitate attackers' lateral movement [29]. Virtualization requires new protection approaches, since conventional physical measures do not guarantee workload isolation or the management of legal liability when the user transfers operational control to the provider [44].

Malware poses a critical threat to IaaS infrastructure, exacerbated by users' inexperience; the redundant use of configuration scripts allows a single vulnerability to propagate across all servers [68]. Non-linear and persistent attack patterns—such as remote reconnaissance and privilege escalation—evade traditional monitoring systems during the transmission and distributed storage phases [43]. Dependence on external certificate authorities and the risk that the cloud administrator may access master keys create inherent distrust that demands certificateless cryptographic schemes [45].

Serverless computing introduces side-channel vulnerabilities, data leakage between functions, and exposure of residual states during container reuse after a cold start [29]. Centralized infrastructure creates single points of failure, while multi-tenancy increases the risk of inconsistencies in information management and privacy breaches [46]. Communication channels face DoS attacks, identity theft, malicious code injection (XSS), and spoofing techniques that compromise the integrity of resources delegated to the cloud [47]. IDS systems cannot scale in ultra-high-speed networks, and AI models have difficulty identifying zero-day attacks; moreover, high computational cost and lack of interpretability limit their real-time implementation [48].

Misconfigurations and storage security flaws are frequent vulnerabilities; organizational distrust of external control represents a critical barrier to service integrity and availability [7]. The convergence of AI and cloud introduces global challenges in which complex cyberattacks can incapacitate critical infrastructures, and the use of the cloud to design synthetic biological agents raises unprecedented biosecurity dilemmas [69]. The high dimensionality of traffic, class imbalance in datasets, and the constant evolution of attack patterns hinder effective intrusion detection in distributed and multi-tenant environments [35].

Cloud infrastructure is vulnerable to internal threats such as flooding, port scanning, and DDoS, which compromise the system's confidentiality, integrity, and availability triad [51]. The multi-tenant model presents critical risks of unauthorized cross-access among clients that share physical or logical infrastructures, requiring operational transparency from the provider (CSP) [50]. Challenges vary by service model: in IaaS, virtualization limitations are the fundamental challenge; in PaaS, interoperability and host sensitivity problems arise; and in SaaS, the focus is on authentication, authorization, and data protection [66]. Attacks by malicious insiders are the most destructive, as such insiders possess legitimate access or elevated privileges that allow them to bypass firewalls and conventional detection systems [70]. The multi-tenant architecture and node heterogeneity create risks of information leakage during synchronization of distributed models, requiring new encryption and anonymization strategies [39]. Operational failures can corrupt or destroy data, leading to authentication problems that block access for legitimate users [36].

VM migration in green clouds is vulnerable to the interception of sensitive data and to compromise of the kernel and hardware state[40]. DDoS attacks saturate bandwidth and server resources, preventing legitimate traffic in infrastructures whose dynamic design breaks conventional perimeter security paradigms [56]. The coexistence of multiple instances on a single physical server allows a vulnerability in a VM to be exploited to escalate privileges and gain unauthorized access to other users' data [11]. Compliance with various international laws such as the GDPR, when information is stored on servers outside the organization's direct authority, represents a fundamental barrier to cloud adoption [55]. The public cloud presents greater security and integrity challenges compared with hybrid or private models, which offer greater control over infrastructure [71].