Search Results (47)

Adversarial machine learning (AML), reinforcement learning (RL), and explainable artificial intelligence (XAI) are increasingly studied as separate problems, yet their interactions under realistic threat conditions remain poorly understood. This review addresses that gap through a systematic analysis of 207 studies selected from 4447 records following the PRISMA 2020 guidelines, covering work published between 2020 and 2026 across cybersecurity and computer vision. A taxonomy of adversarial attacks is constructed across training and inference phases, defense mechanisms are examined with attention to their documented failure modes, and robustness evaluation practices are assessed across the surveyed literature. RL is analyzed in both offensive and defensive roles. Attack agents using RL achieve evasion rates of 74–97% against ML-based detectors, while RL-based defenses report robustness gains of up to 3× over static baselines under comparable threat conditions. XAI receives particular attention because the field treats it almost exclusively as a transparency mechanism, whereas the reviewed evidence shows that it also functions as an attack surface. Attribution methods such as LIME, SHAP, and Grad-CAM produce unreliable explanations under adversarial perturbation, and no system in the reviewed literature certifies that attribution properties are maintained when inputs are manipulated. The review concludes with an analysis of open problems and research directions for building systems that are robust against adaptive adversaries, interpretable under operational constraints, and auditable in environments where AI accountability is a legal requirement. Full article

Deep neural networks (DNNs) have significantly advanced machine perception and reasoning; however, their lack of transparency in decision-making continues to pose a major challenge, particularly in high-stakes domains such as healthcare, finance, and law. This is especially concerning with the black-box nature of convolutional neural networks (CNNs), where the rationale for making a decision can be as important as the decision itself. This paper is driven by a question that is easier to ask than to answer: how can CNNs be made to explain themselves? To answer the question, we wrote a PRISMA-compliant systematic review of 154 studies published between 2017 and 2025. These studies were selected from 4421 studies retrieved through Web of Science, Scopus, IEEE Xplore, and ACM Digital Library. CNN-specific taxonomy was developed. This taxonomy organizes explainable artificial intelligence (XAI) methods on four axes: explanation timing, model dependency, output type, and target component. We found that there is a huge bias in the field regarding post hoc visual methods. Grad-CAM is the most widely cited visual explanation methodology, and within the model-agnostic framework, LIME and SHAP prevail. This research was also the first to analyze standard assessment methods. It was found that out of the 154 studies in the review, 98 used objective methods to evaluate fidelity, stability, or sensitivity. Conversely, fewer than ten of them used human-centered methods to evaluate how tasks were performed, how the users trusted the method, or how the users were prepared to interact with the system. We argue for a dual-reporting convention under which metrics should be reported together at least once, as per the family of metrics. The third contribution is an evidence-based challenge map, where we outline four issues: absence of standardized benchmarks, post hoc mechanism scalability limitations, vulnerability to adversarial perturbations, and the persistent gap between the technical descriptions and human understanding. For each challenge, we propose concrete directions: integrating causal reasoning, adopting participatory evaluation design, and building hybrid transparent architectures. We offer this review as a practical roadmap for researchers and practitioners working toward more explainable deep neural networks. Full article

Connected and Automated Electric Vehicles (CAEVs) are rapidly evolving into complex Cyber-Physical-Social Systems (CPSS), generating structural tensions between technological innovation and public safety. Current research in public governance exhibits significant fragmentation. Scholars frequently isolate data privacy compliance from algorithmic safety auditing, treating them as distinct silos. To bridge this gap, this study applies the PRISMA framework to systematically synthesize 135 core peer-reviewed articles, exposing the endogenous limitations of unidimensional regulatory paradigms. Our analysis yields three central insights. First, traditional “notice-and-consent” models fail under the ubiquitous data collection demands of modern V2X environments. Macro-level policies must translate into foundational Privacy-Enhancing Technologies (PETs) through “Law-as-Code” mechanisms. Second, the opacity of end-to-end algorithmic decision-making deconstructs traditional tort liability systems. This necessitates ex-ante quantitative auditing mechanisms—such as Explainable Artificial Intelligence (XAI) and enhanced Threat Analysis and Risk Assessment (TARA 2.0)—to mitigate adversarial attacks and physical-level safety hazards. Third, overcoming cross-national regulatory fragmentation requires constructing a “dual-track synergistic” governance architecture. This framework institutionalizes the coupling of data lifecycle quality workflows with the algorithmic Safety of the Intended Functionality (SOTIF). Ultimately, this review advocates for adaptive regulatory sandboxes and advances the harmonization and mutual recognition of global standards (e.g., ISO/SAE 21434, UN R155/156). Addressing current methodological and empirical data constraints, future academic inquiry must pivot. Researchers should target the value alignment challenges of Large Language Models (LLMs) in autonomous driving and implement multi-stakeholder participatory policy pilots designed to reconcile diverse social values. Full article

Glaucoma is a leading cause of blindness, and early detection is critical. Convolutional neural networks (CNNs) have shown impressive performance in glaucoma diagnosis, but their black-box nature remains a barrier to clinical use. Existing explainable AI (XAI) methods such as Grad-CAM have limitations in identifying and quantifying subtle regional features. In this study, we propose a method to clarify what CNNs focus on by analyzing how model performance changes under localized adversarial noise. Using VGG16 for glaucoma classification, we applied noise generated by the Fast Gradient Sign Method (FGSM) to the whole fundus image and to specific subregions, then compared the impact on classification performance. Results showed that perturbations to the optic disc, especially its outer margin, had the greatest effect on model performance. This suggests that the CNN captures fine anatomical features such as optic disc cupping and neuroretinal rim thinning, which aligns with what ophthalmologists typically look for. At the same time, perturbations in the macula and perivascular regions also affected performance, indicating gaps between current clinical diagnostic criteria and the CNN’s decision-making process. This approach can help establish the clinical reliability of CNNs and may also reveal features that have not been recognized in conventional clinical practice. Full article

South Africa faces significant challenges in monitoring air pollution from different provinces due to the sparse nature of the sensor network and heterogeneous pollutant sources. Notably, some provinces continue to record a limited amount of data on air pollution, thus making monitoring in those locations problematic. Fortunately, the capabilities of deep learning models to facilitate effective monitoring in data-scarce locations have been highlighted by researchers; however, these models within the context of transfer learning still lack transparency and uncertainty quantification. Using air pollutants and meteorological factors, this study proposes a transfer learning model for particulate matter (PM_2.5) prediction in a data-scarce region. This transfer learning (TL) model leverages an adaptive Bi-directional Gated Recurrent Unit (adaBiGRU) with explainable artificial intelligence (xAI) and uncertainty quantification (UQ) to provide a novel uncertainty-aware adaptation transfer learning (UATL_adaBiGRU) model for a data-scarce location. Variant models based on the adaBiGRU technique, such as the temporal convolution network adaBiGRU (TCN-adaBiGRU) and domain-adversarial neural network adaBiGRU (DANNadaBiGRU), are presented as comparative models. The performance evaluation metrics are root mean squared, R² score and mean squared error. The R² score of pre-trained models in source domain is adaBiGRU (0.888), DANN_adaBiGRU (0.7788) and TCN_adaBiGRU (0.876). Furthermore, other comparative TL models include GRU (0.898), MLP (0.802) and adaptive LSTM (0.886). Afterwards, the pre-trained baseline model (adaBiGRU) was fine-tuned in the target domain dataset and the unpromising result contributed to the proposition of the UATL_adaBiGRU model for a data-scarce location, with R² score of 0.9618. Uncertainty assessment metrics results were also presented for the proposed model. Ablation assessment demonstrates that each component of the UATL_adaBiGRU contributes to enhancing the predictive performance. Again, the Diebold–Mariano (DM) test statistic demonstrates a statistically significant difference between baseline model and UATL_adaBiGRU model. Finally, the local interpretable model-agnostic explanation highlights multi-scaled features as contributing towards the prediction of PM_2.5 in the target domain. In view of this result, model fine-tuning is strongly recommended to enhance the robustness of the proposed uncertainty-aware adaption model in data-limited regions in South Africa. Full article

The following narrative review discusses the use of deep learning and 3D modeling in facial reconstruction from skeletal remains, focusing on accuracy, algorithmic bias, and evidential reliability. Forensic facial reconstruction (FFR) is a multidisciplinary field combining anthropology, medicine, and visual sciences to approximate the facial appearance of unidentified individuals from skeletal remains. Traditional manual methods, based on anatomical knowledge and facial soft tissue thickness (FSTT) measurements, are limited by subjectivity, labor intensity, and inter-expert variability. This narrative review summarizes contemporary AI-assisted approaches, with emphasis on convolutional neural networks (CNNs), generative adversarial networks (GANs), variational autoencoders (VAEs), and diffusion models, which enable probabilistic prediction of facial morphology while accounting for demographic variables such as sex, age, and population ancestry. Key challenges affecting reconstruction accuracy—including dataset limitations, population-specific variability, and algorithmic bias—are discussed, alongside quantitative validation methods and concerns regarding model transparency. Legal and ethical considerations, such as privacy, biometric data protection, and the need for explainable AI (XAI) frameworks, are highlighted. Future perspectives include hybrid expert–AI workflows, the development of globally representative datasets, and the integration of multimodal data sources, including DNA phenotyping, 3D morphometrics, and biomechanical modeling. These advances aim to create standardized, interpretable, and biologically informed frameworks that enable AI to support expert judgment and enhance the reliability of forensic facial reconstructions. Full article

The recent rapid growth of deep learning has significantly reshaped the landscape of computer vision, establishing itself as the preferred paradigm for various tasks. Deep learning methods have demonstrated superior performance compared to previous state-of-the-art machine learning techniques across various fields. This review provides a concise overview of artificial neural networks (ANNs) and some of the most significant deep learning architectures, such as recurrent neural networks (RNNs), generative adversarial networks (GANs) and radial basis function networks (RBFNs). This review not only outlines the historical context and structures of these architectures but also provides a sophisticated understanding of their applications across different computer vision domains. A rigorous and comprehensive overview of these architectures is discussed throughout this review, and an essential systematic comparative analysis based on specific benchmarking criteria is provided. While individual deep learning frameworks excel in distinct domains, selecting the optimal architecture requires a balanced trade-off between algorithmic complexity, computational overhead, data dependencies, and structural interpretability. An intuitive and holistic benchmarking process synthesizes the core characteristics, technical configurations, operational constraints, and developmental pathways toward Explainable AI (XAI) and Green AI sustainability for the examined architectures (ANNs, RNNs, LSTMs, GANs, and RBFNs). Additionally, in this work the advantages and limitations of these architectures are discussed. Furthermore, an investigation of their applications in diverse computer vision tasks is carried out. Full article

As the Internet of Things (IoT) grows, strong, scalable, and adaptive intrusion detection systems (IDS) become increasingly critical for protecting IoT environments. This paper presents a comprehensive and systematic survey of IDS techniques for IoT environments, covering literature from 2021 to early 2026. The review introduces a multidimensional taxonomy that categorizes IDS approaches by detection strategy, learning paradigm, deployment architecture, and evaluation methodology. We examine conventional techniques, such as signature-based and anomaly-based detection, as well as modern machine-learning and deep-learning approaches. Furthermore, emerging paradigms, including Federated Learning, Explainable AI (XAI), TinyML, Large Language Models (LLMs), Transformer, Quantum Machine Learning, Generative Adversarial Networks and Incremental Learning, are analyzed with respect to their applicability to resource-constrained IoT environments. The paper also provides a detailed analysis of publicly available IDS datasets, validation protocols, and evaluation metrics used for benchmarking detection systems. In addition, critical challenges, including dataset realism, adversarial robustness, scalability, privacy preservation, and ethical considerations, are discussed. Finally, we highlight open research directions and propose guidelines for designing next-generation, trustworthy, and scalable IDS frameworks for IoT networks. Full article

Accurately predicting traffic crash severity is critical for intelligent transportation systems, where outcomes emerge from the interaction of infrastructure, environment, traffic control, and human behavior. However, existing approaches face three key challenges: severe class imbalance, computational inefficiency, and limited support for system-level risk understanding. To address these issues, this study proposes a unified and system-aware framework integrating Conditional Tabular Generative Adversarial Network (CTGAN), Tabular Prior-data Fitted Network (TabPFN), and eXplainable Artificial Intelligence (XAI) methods for data augmentation, efficient prediction, and interpretable analysis. CTGAN enhances rare but critical crash states while preserving feature dependencies; TabPFN enables accurate multi-class prediction with limited dataset-specific tuning; and XAI methods quantify the influence of key factors and their interactions. Experiments on a real-world crash dataset from Boston show that the proposed framework achieves competitive predictive performance with less reliance on dataset-specific hyperparameter tuning, while also providing complementary interpretability results from multiple perspectives. The results further reveal that crash severity is jointly shaped by visibility, traffic control, roadside features, and temporal dynamics, highlighting the interconnected nature of risk within the transportation system. By integrating predictive modeling with complementary interpretability analysis, the framework provides a systems-oriented basis for examining how environmental, infrastructural, and temporal conditions jointly relate to crash severity in the studied urban crash data, while offering a methodological reference for broader safety applications that require further validation. Full article

As IoT systems complexity grows, transparent and trustworthy machine-learning intrusion detection systems are crucial. Post hoc explainable AI methods, such as SHAP and LIME, are the most widely used ways to explain how models work, but the degree to which these methods are robust to adversarial conditioning is understudied. In this paper, we propose to create a unified system of evaluating explanation fidelity by using three metrics: sparsity, completeness, and robustness based on minimally distorting DeepFool input perturbations. Our study benchmarks SHAP and LIME across three datasets (BoT-IoT, Edge-IIoT, and N-BaIoT) using four classifiers: CNN, DNN, LSTM, and RF. Our results demonstrate a consistent trade-off: SHAP achieves stronger feature alignment and higher completeness under attack, whereas LIME exhibits greater rank stability in terms of top-k feature overlap. However, LIME also produces more spurious attributions and offers less explanatory power than SHAP, especially in the presence of synthetic features. Our findings reveal that high model accuracy does not guarantee that the provided explanation is also high-fidelity. This investigation highlights the necessity for robustness-aware XAI in cybersecurity and provides reproducible parameters to guide the adoption of XAI in adversarial environments. Full article

Artificial intelligence has evolved to generative AI (GenAI), a paradigm shift that has shifted the emphasis away from the evaluation of existing patterns to the generation of novel biological and medical material. This study examines GenAI achievements in biosciences and medical fields the last five years in these fields using databases such as PubMed and Scopus. The paper highlights the recent evolution in biomedical research from virtual screening to de novo design. It illustrates how models like RFdiffusion and ProteinMPNN leverage “inverse folding” to assemble novel of proteins and drugs. Ultimately, these generative methods yield candidate with enhanced binding affinity and structural stability. For example, exploratory studies suggest GenAI has the potential to address inefficiencies via automatic documentation in the therapeutic sector, and it may enhance research capabilities by using Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) to generate synthetic clinical trial data that preserves confidentiality. In addition, the review argues that though GenAI democratizes medical education through scalable simulations, it raises questions about long-term knowledge retention. Finally, GenAI also offers a transformative “write” capability for biology, but its responsible application will require addressing model “hallucinations” and building Explainable AI (XAI) and robust ethical frameworks. Full article

The rise of the Internet of Medical Things (IoMT) has transformed healthcare through real-time monitoring and improved outcomes but also introduced critical security and privacy challenges. This paper presents a focused survey of Explainable AI (XAI) approaches for intrusion detection in IoMT, emphasizing methods that are lightweight, transparent, and deployable under resource constraints. We first clarify XAI terminology and taxonomy (global vs. local scope; ante hoc vs. post hoc; model-agnostic vs. model-specific) and then systematize recent works from the past five years across cybersecurity sub-domains relevant to eHealth. Representative pipelines span classical ML (e.g., LR, RF, SVM, and XGBoost) and deep models (e.g., DNNs and SRU/LSTM), with post hoc explainers, especially SHAP and LIME, dominating practice on benchmark datasets such as CICIDS2017, NSL-KDD, ToN-IoT, WUSTL-EHMS, and CICIoMT2024. Our comparative analysis highlights consistent gains from model ensembling and interpretable feature selection while uncovering key gaps: limited real-world validation, inconsistent explainability metrics, adversarial brittleness, and the computing cost of explanations at the edge. Full article

Connected autonomous vehicles (CAVs) rely on deep neural network-based perception systems to operate safely in complex driving environments. However, these systems remain vulnerable to adversarial perturbations that can induce misclassification without perceptible changes to human observers. Explainable artificial intelligence (XAI) has been proposed as a potential adversarial detection mechanism by exposing inconsistencies in model attention. This study evaluated the effectiveness of NoiseCAM-based explanation-space detection on the German Traffic Sign Recognition Benchmark (GTSRB) using a single 32 × 32 CNN architecture. Adversarial examples were generated using FGSM under perturbation budgets $ϵ$ = 0.01–0.10, and detection performance was evaluated using accuracy, precision, recall, F1-score, and ROC–AUC. Results show that NoiseCAM achieves detection accuracies between 51.8% and 52.9% with ROC–AUC values of 0.52–0.53, only marginally above random discrimination (0.5). Class-wise analysis further reveals substantial variability in detection reliability across traffic sign categories, with visually structured regulatory signs exhibiting higher separability than complex warning signs. These findings suggest that explanation-space inconsistencies alone provide limited adversarial detection capability in low-resolution, safety-critical perception pipelines. The study contributes to the understanding of the operational limits of explanation-based adversarial detection and highlights the need to integrate XAI signals with complementary robustness or uncertainty-aware mechanisms for reliable deployment in autonomous driving systems. Full article

Cyber adversaries are becoming more sophisticated, creating complex security challenges as digital services expand. The reliability of the firewall is of the utmost importance in the context of network security since it serves as the first line of protection. Penetration testing is an approach used to evaluate the reliability of a firewall and improve security by uncovering exploitable flaws. Frequently, penetration testing solutions are developed using machine learning, and it is of the utmost importance to explain the obtained results during the penetration testing. The emergence of explainable AI (XAI) addresses transparency in ML models, which is essential for informed cybersecurity decisions. Additionally, effective penetration testing reports are crucial for organizations, helping them comprehend and address vulnerabilities with tailored mitigation strategies. This study contributes to firewall security by developing an explainable penetration testing method, which includes two machine learning classification models: a binary model for detecting attacks and a multiclass model for identifying attack types with an explainability feature. This research introduces a novel explainability method that emphasizes significant features related to attack types based on multiclass predictions and proposes an approach using the extended System Security Assurance Ontology (SSAO) to clarify vulnerabilities and suggest alternative mitigation strategies. After evaluating numerous ML algorithms for the CIC-IDS2017 dataset, the Fine Tree model was considered to have the greatest performance. For the binary model, it achieved a validation accuracy of 99.7%, while for the multiclass model, it achieved a validation accuracy of 99.6%. Both models were used to test the firewall for vulnerabilities. Firewall penetration testing using the binary model achieves an accuracy of 82.1%, while the multiclass model achieves an accuracy of 78.7%. Full article

With the increasing complexity of cyber threats and the inefficiency of traditional vulnerability management, artificial intelligence has been increasingly integrated into cybersecurity. This review provides a comprehensive evaluation of AI-powered strategies including machine learning, deep learning, and large language models for identifying cybersecurity vulnerabilities and supporting automated patching. In this review, we conducted a synthesis and appraisal of 29 peer-reviewed studies published between 2019 and 2024. Our results indicate that AI methods substantially improve the precision of detection, scalability, and response speed compared with human-driven and rule-based approaches. We detail the transition from conventional ML categorization to using deep learning for source code analysis and dynamic network detection. Moreover, we identify advanced mitigation strategies such as AI-powered prioritization, neuro-symbolic AI, deep reinforcement learning and the generative abilities of LLMs which are used for automated patch suggestions. To strengthen methodological rigor, this review followed a registered protocol and PRISMA-based study selection, and it reports reproducible database searches (exact queries and search dates) and transparent screening decisions. We additionally assessed the quality and risk of bias of included studies using criteria tailored to AI-driven vulnerability research (dataset transparency, leakage control, evaluation rigor, reproducibility, and external validation), and we used these quality results to contextualize the synthesis. Our critical evaluation indicates that this area remains at an early stage and is characterized by significant gaps. The absence of standard benchmarks, limited generalizability of the models to various domains, and lack of adversarial testing are the obstacles that prevent adoption of these methods in real-world scenarios. Furthermore, the research suggests that the black-box nature of most models poses a serious problem in terms of trust. Thus, XAI is quite pertinent in this context. This paper serves as a thorough guide for the evolution of AI-driven vulnerability management and indicates that next-generation AI systems should not only be more accurate but also transparent, robust, and generalizable. Full article