Can Eyupoglu

When a disaster disables terrestrial infrastructure, the command, control and situational-awareness traffic of the response teams loses its network. Fleets of unmanned aerial vehicles can restore connectivity within minutes, but most designs treat clustering, medium access, energy and failure recovery as separate problems, so a gain obtained in one layer is often lost in another once aircraft start to fail. This paper presents the Resilient Adaptive Clustering and Access Engine (RACE), a cross-layer control loop for multi-UAV emergency networks in the beyond-5G setting. At every control epoch, RACE places cluster heads with a greedy rule that maximizes the end-to-end availability offered to the fleet while accounting for residual energy, associates members with the head that serves them best without overloading any cluster and protects an availability-responsive share of capacity for mission-control traffic. The evaluation uses a matched-scenario Monte Carlo model with correlated mobility, air-to-air and air-to-ground link availability, co-channel interference, energy depletion, degraded propagation, prioritized traffic and injected aircraft losses, and it charges the signaling, failure-detection delay and reconfiguration transients of the control loop. RACE is compared with a flat multichannel access profile and with three clustering schemes from the literature, both as designed and with every scheme given the same control allocation. For dense fleets, RACE delivers the most traffic and the widest robust coverage and keeps mission-control delivery near 0.9 across all failure bursts. The protected share removes the long control-delay tail of any clustering rule, and with the same share the placement and association rules of RACE still add traffic, coverage and a shorter estimated delay tail, with a margin that grows as the network becomes more fragile. Ablations, a long-mission energy study and sensitivity sweeps confirm that these conclusions are stable.

Hybrid automatic repeat request (HARQ) remains the backbone of link-layer reliability, yet it was designed for a setting in which every bit matters equally and the channel of the next slot is unknown. Neither assumption holds in semantic communication over time-varying channels: some parts of a message are decisive for the receiver’s task while others are not, and the deep fade that destroyed a first transmission is usually still present when an immediate retransmission arrives. This article proposes predictive semantic-aware HARQ (PSA-HARQ), a retransmission framework that makes its decisions at the level of individual semantic units and bases them jointly on how much each unit matters and on how the channel is expected to evolve. Semantic importance is extracted from a Transformer semantic encoder, a lightweight recurrent predictor forecasts the channel a few slots ahead, and a finite-horizon optimal-stopping controller decides, for every unit and every slot, whether to transmit, wait for a better slot, or abandon the unit. The forecast enters the controller through a confidence-weighted estimator whose weight is tracked at run time from the observed prediction errors, so that the policy falls back to the statistics-only optimum whenever the predictor becomes uninformative. Simulations over a vehicular situational-awareness task and over a real public record-classification dataset, with finite-blocklength coding and time-correlated Rayleigh fading, show that PSA-HARQ matches or exceeds the task accuracy of conventional HARQ in the low-SNR regime where retransmissions are most frequent, while transmitting far fewer blocks and consuming far less energy at the same channel condition. A cost-matched ablation shows that semantic importance and channel prediction contribute complementary gains. The results indicate that anticipating the channel, rather than merely reacting to it, is a practical and largely untapped lever for reliable and efficient 6G semantic communication.

Adaptive semantic communication can save radio resources by matching the representation rate to channel quality, yet its decisions become unreliable when channel state information reaches the transmitter with delay. This study examines that overlooked failure mode and develops a prediction-aided framework for rate-adaptive transmission of Internet of Things sensor streams. A compact long short-term memory network forecasts a distribution of the receive signal-to-noise ratio (SNR), and a risk-aware controller selects the smallest rate whose expected semantic distortion under that forecast meets a fidelity target. The controller is independent of the semantic codec and accounts for forecast uncertainty rather than relying on a single predicted value. The framework is evaluated on six real sensor streams, over time-correlated fading channels and on SNR logs measured in a commercial 5G/LTE network, with all schemes held to the same average bandwidth and with confidence intervals over independent channel realizations. Adaptation based on stale channel information loses its advantage over fixed-rate transmission as mobility or feedback delay increases, both in simulation and on measured logs. Channel prediction recovers a substantial part of this loss, a single predictor serves the whole mobility range without retraining, and the learned forecast remains reliable when the channel statistics assumed by a model-based predictor are wrong. Prediction also keeps its advantage over stale channel information under drifting mobility, varying channel-estimation quality and frequency-selective multipath, and a predictor trained for one mobility regime can be adapted to another within seconds. A second-order analysis, whose assumptions are verified on the trained codec, explains how forecast uncertainty and the curvature of the codec’s distortion response affect the rate decision, including how often uncertainty changes the choice among discrete rates. The resulting framework is lightweight and suitable for practical semantic links operating with delayed feedback.

Dyeing techniques have a long history dating back to the Bronze Age and are among the oldest artistic and technological activities of humanity. In the early periods, plant-based dyes were applied d...

Despite the strong capabilities of Large Language Models (LLMs) across a wide range of natural language processing tasks, their outputs can still be unreliable. Hallucination is widely used as an umbrella label for several related failures, but its operational meaning varies across tasks and studies. This work provides a comprehensive review of recent advances in understanding, detecting, and mitigating hallucinations in LLMs. We first distinguish hallucination from the separate evaluation dimensions of factuality, faithfulness, and internal consistency. We then introduce an author-proposed organisational taxonomy covering manifestations commonly discussed under hallucination, including factual contradictions, context conflicts, fabrication, self-contradiction, and emerging failures in reasoning traces and agentic trajectories. We review detection methods based on level of access to the model, ranging from white-box approaches to grey-box uncertainty estimation and black-box consistency and verification methods. We further organize these approaches according to the signals they exploit, including uncertainty, semantic consistency, external evidence, and internal representations, and discuss their strengths and limitations across different deployment settings. Finally, we survey mitigation strategies across the LLM lifecycle, including (1) data-centric approaches such as data curation and retrieval-augmented pre-training, (2) model-centric approaches such as supervised fine-tuning, preference optimization, abstention-aware reinforcement learning, and knowledge editing, and (3) inference-time approaches such as retrieval-augmented generation, self-verification, and decoding or activation-level interventions. Rather than viewing these strategies as isolated solutions, we argue that reliable hallucination mitigation requires complementary controls across multiple stages of the model lifecycle. Key challenges include scalable and provenance-aware data curation, the alignment–capability trade-off, reliable composition of mitigation techniques, and the need to evaluate and modify reasoning and agentic processes rather than only isolated factual associations.

Background/Objectives: Accurate segmentation of histopathological images while preserving cellular morphology in computer-aided diagnostic systems is critically important for the diagnosis and staging of colon cancer. However, conventional metaheuristic algorithms performing multi-level thresholding on such complex tissues often suffer from premature convergence by becoming trapped in local optima as the search space increases. To address this limitation, this study proposes a new label-independent hybrid optimization algorithm focused on colon adenocarcinoma segmentation. Methods: The proposed algorithm hybridizes the global exploration capability of the Artificial Rabbit Optimization (ARO) algorithm with the local exploitation ability of the Ant Lion Optimization (ALO) algorithm through an optimized fixed transition ratio, thereby enabling efficient localization of cellular density valleys. Results: The principal findings obtained from the LC25000 colon cancer dataset demonstrate that the ARO-ALO algorithm achieves stable performance with high SSIM (0.8043) and FSIM (0.8595) scores while preserving the histopathological hierarchy. Furthermore, the preservation of diagnostic morphology after segmentation is statistically validated by the high Pearson (0.9870) and Spearman (0.9948) correlation coefficients. In addition, supplementary generalization experiments are conducted on the Oral Squamous Cell Carcinoma (OSCC) and pulmonary circulation vessels datasets to verify the tissue-agnostic nature of the algorithm. Conclusions: Consequently, the ARO-ALO algorithm emerges as an efficient alternative for clinical decision support systems.

Developing a privacy-preserving data publishing algorithm that prevents individuals’ identity information from being disclosed without ignoring the utility of the data is still an important goal to be achieved. Finding the optimal balance between data utility and data privacy is an NP-hard problem. In this study, a new k-anonymization-based data anonymization algorithm is proposed. The proposed algorithm partitions the data space with a new efficient partitioning strategy based on the k-dimensional tree (KD-tree), resolves the boundary problem while maintaining the trade-off balance, and introduces a new mechanism to address the problems caused by outliers. Moreover, it can be applied to both numerical and categorical data. The experimental results indicate that the proposed algorithm achieves competitive or improved performance compared with the baseline algorithms across seven commonly used evaluation metrics. Overall, the findings suggest that the proposed framework can improve the utility–privacy trade-off in multidimensional k-anonymization.

Background/Objectives: Privacy preservation in big data environments is an NP-hard optimization task that requires the satisfaction of k-anonymity and l-diversity constraints to ensure data utility. Methods: This study proposes a novel hybrid optimization approach, adaptive hybrid FOX–RUN Intensive Mutation (PA-FRIM), to address the privacy–utility trade-off in anonymization process. The proposed approach integrates FOX-based global exploration with RUN-based local search using a hybrid adaptive control strategy and intensive mutation search to improve solution diversity in highly constrained solution spaces. Results: The experimental study on the Adult and Bank Marketing datasets shows that PA-FRIM exhibits stable convergence behavior compared to competing methods. The results indicate that full privacy is achieved on the Adult dataset with a violation value of 0.00, and information loss is minimized with an NIL measure of 0.5686. From the analytical utility perspective, PA-FRIM ensures data usability, even in the constrained region, achieving classification accuracies of 89.61% on the Bank Marketing dataset and 84.90% on the Adult dataset. Conclusions: By using a multi-metric evaluation strategy, PA-FRIM provides a robust optimization framework that eliminates privacy violations while maintaining high analytical performance.

The secure management of classified documents containing sensitive information is critical for governments, military organizations, and the industry. Traditional data loss prevention (DLP) systems lack robustness against insider threats, particularly regarding access log integrity and tamper-proof auditing. To address log security, the previous literature has proposed multiple solutions, including private and hybrid blockchain models (e.g., Ethereum + MultiChain) to ensure audit trail integrity. However, hybrid architectures often face challenges such as unpredictable transaction costs (gas fees) and potential privacy risks when scaled for enterprise DLP logs. Conversely, private architectures may require higher resources, potentially causing bottlenecks on endpoints. In this paper, we propose an optimized Blockchain-Based Classified Document Management System (B2CDMS) utilizing a permissioned architecture. Our work demonstrates the challenges, advantages, and weak points of current solutions. We optimized a permissioned blockchain (BC) (Hyperledger Fabric v2.5) with an External Chaincode Builder using the Chaincode-as-a-Service (CCaaS) pattern. We compared our proposed private architecture with a hybrid architecture (Ethereum + MultiChain) and a public solution (Ethereum). We conducted a comprehensive analysis using pseudo Trellix ePolicy Orchestrator (ePO) Data Loss Prevention (DLP) logs. Experimental results on an Apple Silicon M4 (Apple Inc., Cupertino, CA, USA) testbed show that the proposed architecture achieves a throughput of 845.8 Transactions Per Second (TPS) with a sub-second latency of 55 ms, aiming to eliminate the bottlenecks of public blockchains. Furthermore, the system introduces a privacy-preserving hashing mechanism (i.e., committing only deterministic Secure Hash Algorithm 256-bit (SHA-256) digests to the immutable ledger while keeping the actual sensitive Personally Identifiable Information (PII) strictly in off-chain databases) compliant with General Data Protection Regulation (GDPR). It ensures that classified document metadata remains immutable and secure against rogue access benefiting from admin privileges. This study concludes that permissioned blockchain architectures offer a scalable and resource-efficient solution for forensic evidence preservation throughout the classified document lifecycle.

Nowadays, artificial intelligence (AI) has been widely used in many fields, such as fine arts and design, education, accounting, media, and judicial systems. AI provides significant potential benefits such as the ability to produce similar products in short periods of time, automate processes, increase efficiency, and provide data-driven analyses. However, this rapidly developing use of AI brings with it a number of complex ethical issues and risks. These issues include lack of transparency, accountability uncertainties, data privacy and security breaches, algorithmic bias and discrimination, copyright and copyright ownership issues, impacts on the labour market and employment concerns, potential for manipulation and disinformation, reduced human touch, and unpredictability of systems. This situation emphasises the importance of the concept of “responsible artificial intelligence”, which is an approach sensitive to ethical values, security, transparency, and user rights in the design, development, implementation, and use processes of AI. Responsible AI includes fundamental principles such as fairness, transparency, accountability, robustness and security, data governance, laws and regulations, human oversight, and societal welfare. Overcoming these ethical challenges requires collaboration among various stakeholders, such as artists and designers, technologists, decision-makers, lawyers and educators, and the implementation of appropriate policy and legal regulations. This study aims to reveal the importance and necessity of responsible and ethical use of AI by examining the ethical risks and issues in different application areas of AI through a systematic approach through relevant literature and real examples. In order to maximise the societal benefits of AI, it is essential to carefully address ethical dimensions and develop continuously updated ethical standards.