Machine learning (ML)-based Network Intrusion Detection Systems (NIDS) remain vulnerable to adversarial perturbations. This study evaluates fixed-mask FGSM and PGD attacks on NSL-KDD using six clean baselines, black-box transferability, category-level analysis, and PGD-based adversarial training. Constrained and unconstrained attacks produced the same aggregate evasion rate at ε=0.1(28.9%), while PGD caused a systematic paired shift relative to clean detection (McNemar χ²=2501.0, p<0.001). Transfer evasion ranged from 0.7% to 14.8%. Category-level evasion ranged from 14.6% for DoS to 100% for R2L/U2R, together with a large zero-perturbation subgroup whose cause remains unresolved. On model-specific eligible sets at ε=0.1, PGD evasion was 28.9% for the undefended model and 7.9% for the defended model. A direct paired common-set comparison using independently generated model-specific PGD attacks also favored the defended model (18.2% vs. 4.4%, McNemar χ²=1012.68, p<0.001), with a 5.1-percentage-point clean-accuracy cost. Under the adopted feature-space setting, freezing categorical and discrete indicator fields did not reduce aggregate evasion, while adversarial training produced a measurable robustness-accuracy trade-off.

Cloud and edge infrastructures now run the data and control loops of healthcare, smart cities, industry and finance, so one compromised node can disrupt an entire digital ecosystem. Protecting them requires intrusion detection that is accurate, privacy-preserving, resistant to manipulation and explainable. This paper addresses federated intrusion detection across cloud-coordinated edge sites, some of which may be compromised and poison the shared model. Existing federated detectors accept model updates on trust, while Byzantine-robust aggregation rules lose accuracy under the heterogeneous traffic of edge sites. Moreover, explainable AI serves mainly post-hoc analyst reports, and zero-trust access control and blockchain logging are seldom linked to learning. We propose XTrust-CE, whose core, Explanation-Fingerprint Trust Aggregation (EFTA), verifies each client update through the class-wise attribution fingerprint of the updated model on a verifier-owned reference set and through its reference behavior, admits only verified, clipped and reputation-weighted updates, and anchors trust scores in a PBFT-based ledger used by a risk-adaptive zero-trust policy. To the best of our knowledge, this is the first framework to use explanation fingerprints for federated update verification and couples that trust with zero-trust enforcement and tamper-evident logging. In a 20-site emulated testbed, EFTA preserved clean accuracy (96.71% versus 96.85% for FedAvg) and, with 40% poisoned sites, kept a 97.06% detection rate compared to 88.28% for Median and 85.03% for Trimmed Mean. Under targeted label flipping it quarantined every poisoned update, holding the ransomware and exfiltration miss rate at 3.52% (FedAvg 27.25%), while the zero-trust policy cut allowed evasive attack flows from 87.4% to 33.7%.

The use of Large Language Models (LLMs) in automated software testing and debugging is a growing area of interest. This study is an experimental investigation of the effectiveness of LLM prompting techniques for software fault detection, fault localization, and automated bug repair using mutation testing. A total of 159 lines of code and 33 test cases were evaluated in four Python programs. 32 mutants were created with six mutation operators, and 28 mutants were killed, resulting in an overall mutation score of 87.5%. Three prompting strategies (zero-shot, few-shot, and iterative test-feedback) were compared based on Fault Detection Rate (FDR), Fault Localization Accuracy (FLA), Repair Success Rate (RSR), Precision, Recall, and F1-score. The few-shot prompting and the zero-shot prompting methods showed balanced detection performance with the highest FDR (92.9%) and F1-score (91.2%) respectively. The experimental findings showed that there was a significant difference between the performance of fault detection and fault debugging. The prompting strategies yielded high fault-detection performance, but their effectiveness was significantly reduced for fault localization. The results suggest that the evaluated LLM prompting techniques were more successful in detecting incorrect behavior than in accurately determining the specific code locations that are responsible for the behavior or in creating validated corrective changes.

Smart City (SC) refers to an innovative idea that makes use of Information and Communication Technology to organize cities and create digitally connected urban environments. The extensive use of Internet of Things (IoT) devices and cutting-edge technologies makes network security in smart cities essential. Establishing and executing strong security procedures often requires public-private collaborations. To improve the privacy and security of smart city networks, various network protocols and systems have emerged beyond conventional measures. However, these solutions face significant limitations, including high algorithmic complexity, longer processing times, reduced maintainability, and a lack of robust authentication and verification mechanisms. To overcome these drawbacks, a Deep Quantile Regression Neural Network (DQRNN) with a multifactor graphical authentication system is developed for improving smart city security. Initially, in smart cities, data is collected via sensors and transmitted to a cloud server for centralized storage. To access data from the cloud server, a Multifactor Graphical-based Authentication Scheme (MGAS) is used, which safeguards data integrity and prevents unauthorized access. For intrusion detection during transmission from the user to the cloud, features of the transmitted data packets are collected and subjected to various processing steps. Initially, the data undergoes pre-processing using Aggregation-based Group Normalization (AGN) and an Edge Generation Graph Auto-Encoder (EGGAE) to improve data quality. Feature extraction then follows using Regularization-based Locality Sensitive Discriminant Analysis (RLSDA) to isolate the features that are most relevant for intrusion detection. The data is then classified using a DQRNN tuned with the Giant Armadillo Optimization Algorithm (GAOA) to detect intruders. The proposed model achieves an accuracy of 96.1%, a selectivity of 96.5%, and a precision of 96.4%; the execution time for the MGAS is 1.2 s and the security level is 93.8%. Consequently, the proposed methods are highly suitable for real-time smart city applications, enhancing network security through robust authentication and intrusion detection.

The Internet of Medical Things (IoMT) is reshaping continuous patient monitoring, yet the authentication and key-agreement (AKA) mechanisms protecting medical telemetry remain exposed to two threats that current blockchain-based schemes do not jointly address. Protocols secured with elliptic-curve cryptography offer no protection against a future quantum adversary, exposing lifelong health records to harvest-now-decrypt-later attacks; and schemes claiming device-capture resistance frequently store long-term secrets in device memory, so a captured node still leaks the values needed to forge sessions. This paper presents QP-BAuKAP, a quantum-resistant, physically-unclonable-function (PUF)-anchored three-party AKA scheme for a three-tier IoMT architecture of device, hospital server, and cloud server, built on a permissioned Hyperledger Fabric blockchain. QP-BAuKAP replaces device-stored secrets with a PUF response reconstructed at run time through a fuzzy extractor, derives the session key from a hybrid of an ephemeral ECC contribution and an ML-KEM (Module-LWE) encapsulation so that confidentiality holds as long as either the elliptic-curve or the lattice problem remains hard, and adds an on-chain Merkle revocation registry for dynamic device decommissioning. We prove semantic security in the Real-or-Random model through a sequence of games reduced to the ECDDH assumption, the IND-CCA security of the KEM, and PUF unpredictability. We corroborate this with Scyther and ProVerif under the Dolev-Yao model. A performance study on a Fabric v2.5 network benchmarked with Hyperledger Caliper shows that QP-BAuKAP achieves quantum resistance and genuine device-capture resistance at modest overhead and remains practical for constrained IoMT deployments.

The Ethereum transaction network contains legitimate transactions as well as fraudulent transactions related to phishing, fraud, and other scams. Although traditional detection models rely on transaction-level features, relationships between senders, receivers and transactions may provide additional information for the detection of malicious intent. However, controlled evaluations comparing tabular and graph-based features in binary and multiclass classifications are still limited. To this end, this paper presents a hybrid tabular-graph framework for fraud detection at the Ethereum transaction level. The proposed methodology was assessed against 71,249 transactions recorded between October 2017 and December 2019, including 57,000 regular transactions, 3,110 phishing transactions and 11,139 fraudulent transactions. From the sender and the recipient addresses, a heterogeneous client-transaction-merchant graph is generated. The structural representation is then extracted using Graphical Convolutional Networks (GCN), Graph SAGE, and Fast Inductive Graph Representation Learning (FI-GRL), and combined with selected tabular attributes. The ensemble classifiers such as Random Forest, XGBoost and LightGBM are used for classification. The proposed approach is assessed using a 5-fold stratified cross-validation without re-sampling for binary classification (normal vs. fraudulent) and multi-class classification (normal, phishing and fraud). The results show that the augmenting graph structural features improve the classification. The FI-GRL and LightGBM combinations provided the highest total performance, with Mathew Correlation Coefficient (MCC) of 0.9912 and 0.9894 respectively on binary and multiclass classification compared to LightGBM with MCC of 0.8373 and 0.8124 respectively for binary and multiclass classification with tabular features. These findings suggest that the integration of structured information derived from graphs, in addition to traditional table attributes, can significantly improve the detection of transaction level fraud in Ethereum network.