Computer Science and Mathematics

Abstract: Hardware noise and finite sampling perturb the fidelity estimates forming a quantum-kernel Gram matrix. We measured how far three hardware-reconstructed Gram matrices depart from an exact statevector reference for one frozen four-qubit ZZ feature map on N = 24 indoor air-quality windows, executed on ibm_fez at 1024 shots per circuit in three single, non-interleaved jobs: baseline, dynamical decoupling alone, and gate twirling alone. All were complete, finite, and positive-semidefinite. Off-diagonal root-mean-squared error (RMSE) against the reference was 0.0878, 0.0864, and 0.0427; full-matrix centered kernel alignment (CKA) ranged 0.933–0.989 and the post hoc diagonal-excluded (U-centered) CKA 0.816–0.986. The gate-twirled job deviated least on every reported geometry axis; its baseline contrasts are deletion-stable for the Spearman, mean-absolute-error, RMSE, and full-matrix CKA diagnostics, while the Pearson and diagonal-excluded contrasts fall just below that convention. Dynamical decoupling was not separated from the baseline. The observed error exceeded both finite-shot reference scales, so, under those sampling-only models, sampling does not explain it. Centered kernel–target alignment did not track reconstruction fidelity and stayed at or below each label-permutation reference: implementation fidelity and task relevance are distinct diagnostic axes. All configuration-level statements describe three realized jobs on one backend; no mitigation-efficacy, classifier-superiority, forecasting, or quantum-advantage claim is made.

Abstract: In recent years, research in hybrid cryptography has provided significant advancements for the world of applications in cybersecurity, however, implementing suitable hybrid cryptographic schemes for IoT environments is still an open research issue. To address the need for a hybrid Key Encapsulation Mechanisms (KEM) for constrained IoT environments, this paper introduces five variants of SEASON KEM, based on the Quantum Superiority Fighter construction. The empirical benchmark demonstrates a significant improvement over state-of-the-art Hybrid KEM, achieving a 20% speed improvement in KeyGeneration and Decapsulation, a 14% improvement in Encapsulation, and additional improvements in memory overhead. The results prove the efficacy of lightweight post-quantum cryptography, and underscore the importance of research in resource-constrained post-quantum security.

Abstract: Applications routinely treat the names, paths, and extensions carried inside archive manifests and download metadata as though they had chosen those values themselves, when in fact an attacker did. A tar entry pathname, an HLS subtitle URI, and a shortcut extension are all attacker-controlled strings, yet each is later consumed verbatim as a filesystem decision. We study three 2026 CVEs that share this structure. CVE-2026-63134 (CISA Malcolm, CVSS 5.4) is a directory traversal in which directory entries skip the path guard that file entries pass, so os.path.join(dest, "/tmp/x") evaluates to /tmp/x (Python discards every argument before an absolute component) and the destination silently vanishes. CVE-2026-63133 (CVSS 6.5) is an inode-exhaustion denial of service in the same extractor, where an archive of 50,000 empty directory entries, about 235 KB compressed, creates 50,000 filesystem objects at an amplification we reproduce near 0.21 objects per compressed byte, with a depth-axis variant that yields many objects from a single entry. CVE-2026-50023 (yt-dlp, CVSS 8.3) is executable-shortcut injection, in which a permission scoped to the -write-link feature was hoisted into a global extension allowlist, so a subtitle track whose URI ends in .desktop writes an executable shortcut under -write-subs. The unifying class is the trust an application places in attacker-controlled filesystem metadata, and within it we identify two failure modes: an inconsistently applied guard, where a security predicate runs on one representation or context but is omitted at a semantically equivalent sink, which accounts for the traversal and the shortcut cases, and a missing invariant, where no resource budget is enforced anywhere, which accounts for the inode case. For each we give the exact code, working payloads, and amplification figures reproduced on a documented environment, together with three static audit rules evaluated on the real vulnerable and patched source, one precise and two heuristic.

Abstract: Large language models (LLMs) have achieved remarkable performance, prompting the development of LLM-based agents to tackle complex, open-world tasks. Despite these advances, autonomous agents still face critical bottlenecks in environmental perception, multi-step reasoning, and long-term opera-tional consistency. Graphs provide a systematic solution to these challenges by explicitly modeling entities, relationships, and dependencies, thereby establishing a rigorous foundation for reasoning and interaction. While recent literature explores this integration, existing surveys primarily focus on operational workflows and architectures, lacking a systematic analysis of how graphs fundamentally elevate agent intelligence. To bridge this gap, this survey presents a comprehensive review of Graph-Augmented LLM-based Agents (GLAs). We systematically decompose agent architecture to distill five core capabilities required for advanced intelligence: perception, reasoning, memorization, generaliza-tion, and interaction. We explicitly address why graphs are uniquely suited to bolster these capabilities and how they are practically integrated, proposing a novel taxonomy that organizes existing studies along these five dimensions. Furthermore, we review representative real-world applications of GLAs across diverse domains and outline key challenges and promising directions for future research. A curated repository containing the literature reviewed in this survey, along with the latest advancements in this field, is maintained and regularly updated at https://github.com/sqs17/GLA.

Abstract: Energy constraints along with scalability and vulnerability to several malicious attacks including Blackhole, Grayhole, Flooding and Scheduling attacks, are serious impediments for wireless sensors networking deployment. Therefore, we present FlowerSec, a novel topology design for wireless sensor networks along with an AI enhanced security scheme which has the ability to overcome afore stated downsides by design. The sensor network is configured on a hierarchical topology of “Flowers,” each comprising a central, resource-rich “Seed” node (Cluster Head) and up to four “Leaf” nodes. The security boundaries that emerge naturally within such topology along with the ease of operating the energy efficient sleep mode, has given us use to propose a scheme which has high-security guarantees with minimal energy loss. Hence a lightweight supervised AI based Intrusion detection System is deployed at the Seed node. Our LightGBM based model with SHAP feature selection that selects only the top 8 network features can obtain an optimal Macro F1-score of 0.9494 on WSN-DS real world dataset (number of instances are 374,661). The classification at CH occurs at 0.0066 ms per sample proving this is a computationally cheap method and can be performed at resource-constrained nodes such as CHs. Beside this, the supervised engine is complemented with an unsupervised AutoEncoder which is trained on pure normal traffic and able to isolate the benign traffic with reconstruction loss difference of >300 times (AUC-ROC score is 0.9819). We adopt confidence gated escalation protocol which sends the packets having low confidence scores of the supervised engine for detection of novel unseen attacks at a small bounded energy cost of the unsupervised engine.

Abstract: Real-time OS has strict time constraints, and violation of time constraints may lead to serious accidents depending on the system in which it is installed. In addition, real-time OS schedules each task to improve CPU execution efficiency. Therefore, it is important to justify the scheduling of tasks in order to keep the time constraints of the system. Exclusion control between tasks using semaphores may cause [1] of task priority. When priority inversion occurs, the processing of a low-priority task is prioritized over a high-priority task, which can easily lead to a time-constraint error. In this paper, we conducted case studies of a new verification for task management modules of real-time OS using a bounded model checker CBMC [3] based on SAT solver with full C/C++ support and a bounded model checker ESBMC [10] based on SMT solver. Through case studies, we were able to formally verify the validity of the task scheduling process and the occurrence of priority inversion. This guarantees the validity of the scheduling of the system and is effective in finding realistic time constraint errors. Real-time OS is ASP3 kernel based on the ITRON [4] specification published by TOPPERS (Toyohashi OPen Platform for Embedded Software) [2].

Abstract: Execution-based verification has been shown to be effective in enhancing the mathematical reasoning abilities of large language models due to its computational soundness guarantees and dependency-aware filtering. Previous works involving preference optimization often include reward models that utilize Bradley-Terry assumptions, which fail to capture the logical dependencies and execution consistency requirements essential for scientific and computational reasoning tasks. In this paper, we introduce a novel method for generating computationally sound solutions accompanied with corresponding dependency graphs for execution-consistent preference optimization. Our approach begins with the construction of a high-quality scientific reasoning dataset by incorporating UltraFeedback prompts, base model generations, computational verification, and execution consistency results. Next, we construct dependency graphs by extracting reasoning step expressions, the computational prerequisites needed for the expressions, and the derivability relationships of the expressions from the previously collected dataset. Based on this extracted information, we generate corresponding execution consistency scores to accurately capture the mathematical verification process. Appending the generated execution consistency scores to each reasoning step results in data consisting of paired filtered reasoning steps and their corresponding execution consistency scores. Training Llama-3-8B and DeepSeekMath-7B with this corpus achieves substantial improvements across scientific reasoning domains: +17.0\% on MATH, +15.1\% on GSM8K, while extending our Scientific Feasibility Control framework to achieve 50.1\% accuracy on PhyX multimodal physics reasoning—outperforming DeepSeek-R1 (49.8\%) and OpenAI o3-mini (48.2\%)—with 91.7\% scientific validity coverage at \(\alpha = 0.10\) confidence level and 73\% reduction in scientific law violations across architectures, leading to the creation of the CCPO family of models.

Abstract: Large language model (LLM) agents are evolving from single-turn text generators into long-running systems capable of multi-step reasoning, tool use, persistent memory, and reusable skill development. Reinforcement learning from verifiable rewards (RLVR) can improve agent reliability, but existing approaches commonly treat verification as a fixed component of the reward structure, creating challenges related to sparse credit assignment, verifier reliability, reward hacking, path compliance, and verification cost. We introduce Budgeted Adaptive Verification with Asymmetric Rewards (BAVAR), a framework that formulates verification as a sequential, resource-constrained decision problem. BAVAR selectively determines what to verify, when verification is valuable, which verifier to invoke, and how verified evidence should influence learning based on uncertainty, action criticality, verifier reliability, expected verification value, and remaining computational budget. It combines reliability-gated positive process rewards with persistent penalties for path violations and extends verification to persistent memories and reusable skills. In an illustrative evaluation under matched verification budgets, BAVAR achieved 72.6% safe verified success, compared with 67.1% for uniform dense verification and 58.4% for outcome-only RLVR. Relative to dense verification, BAVAR reduced verification cost per safe success by 45.5% while using 47.8% fewer verifier tokens. It also reduced verified path violations to 7.2 per 100 trajectories, compared with 20.4 for outcome-only RLVR, and reduced successful verifier exploitation to 4.8%, compared with 15.9% for a single-judge baseline and 18.6% for outcome-only RLVR. For persistent artifacts, BAVAR achieved 88.7% held-out reuse success, compared with 81.3% for dense verification. These simulated findings suggest that adaptive allocation of verification resources can improve safe agent performance while reducing computational and verification overhead, providing a principled foundation for scalable verification of long-running autonomous agents.

Abstract: Searching for the ground state of an Ising model is a fundamental NP-hard problem and serves as a common formulation for a wide range of combinatorial optimization problems. In the absence of external fields, the Ising model is equivalent to the MAXCUT problem, where spins correspond to vertices and interaction coefficients correspond to edge weights. In this paper, we propose a framework for generating large-scale MAXCUT (Ising model) benchmark instances with certified optimal solutions by leveraging Not-All-Equal 3SAT (NAE3SAT) and state-of-the-art SAT solving technology. Random and community-structured NAE3SAT instances are generated near the satisfiability phase-transition region and solved exactly using the SAT solver Kissat. The resulting formulas are then converted into equivalent MAXCUT instances, enabling direct evaluation of approximation algorithms against known optimal solutions. Using the proposed framework, we generate benchmark instances containing up to approximately 100,000 vertices. Experimental evaluation using Optimized Simulated Annealing (OSA) shows that near-optimal solutions can be obtained consistently, achieving approximation ratios above 99.8\% even for the largest instances. We additionally investigate an equation-planting approach for NAE3SAT and observe that the resulting planted instances are substantially easier than the random and community-structured instances. The proposed benchmark generation framework provides large-scale MAXCUT instances with certified optima and offers a bridge between SAT solving and Ising-based optimization, enabling direct comparison between state-of-the-art SAT solvers and MAXCUT solvers on related problem instances.

Abstract: Deepfake detection is usually measured on curated benchmarks, yet synthetic video often reaches audiences through social-media platforms that compress, resize, re-encode, crop, and repost it before downstream detection. The practical question is therefore not whether a detector separates real from fake under laboratory conditions, but whether the evidence it depends on still exists at the moment a decision must be made. This survey reviews Ncorpus works on generative audio-video deepfakes across the full lifecycle of creation, platform distribution, detection, provenance, and remediation. We organize the review around five forensic assumptions that detectors implicitly rely on: that a manipulated region leaves a boundary, that a generator leaves a stable fingerprint, that test data resemble training data, that forensic signals survive processing, and that low-level clues are sufficient. Using these assumptions as a common lens, we present an evidence-based taxonomy of generation methods, review signal-driven, learning-based, reasoning-based, agentic, and adversarially robust detection, and examine provenance mechanisms including content credentials and watermarking alongside emerging disclosure and takedown regulation. Two findings recur. First, foundation-model generation can weaken several assumptions at once, since fully synthetic and jointly generated audio-video content may remove compositing boundaries and reduce exploitable cross-modal inconsistencies. Second, the literature is concentrated on creation and detection, with distribution, provenance, and remediation together accounting for PctTail of the corpus. We close with evaluation protocols that report which assumptions a benchmark actually exercises. Project Page: https://vectorinstitute.github.io/deepfakes-survey-2026/

Abstract: Persistent coding-agent work can leave concurrent input requests, completed results, blockers, and quiet unfinished tasks. AiWingman is an open-source macOS menu-bar companion that derives continuity cues from locally retained Codex task evidence. The dashboard and optional Wingman review are separate. Obsolete or abandoned labels require reversible user confirmation; silence and age are insufficient. This article describes the system but evaluates only its pure continuity-policy layer. The evaluation excludes both readers, the interface, graph signals, token accounting, Codex compatibility, and remote-review quality. It was retrospective and same-project: the Swift policy predated the written specification and 155 expected-output fixtures. The post-implementation specification and fixtures were frozen before the archived run and were not preregistered. They may reflect implementation knowledge, and repository chronology does not exclude earlier exploratory or unarchived runs. A separately implemented same-project Python oracle generated the expected outputs from the written rules. The Swift runner matched 155/155 fixtures. For each fixture, evaluations 2-100 matched evaluation 1 (15,500 evaluations; 15,345 nontrivial comparisons). A post-freeze supplementary driver launched 100 fresh processes on one arm64 Mac. All processes returned the same corpus digest. Across 70 unconfirmed lifecycle fixtures, none produced a confirmed-obsolete or confirmed-abandoned state. Later exact-commit continuous-integration (CI) jobs reproduced both the conformance result and digest on hosted arm64 and native x86_64 macOS runners. The findings establish exact agreement with the 155 frozen, same-project, specification-derived fixture outputs. They do not establish general specification correctness, external validation, human benefit, superiority, exact token cost or waste, security isolation, broad determinism, macOS 13 runtime compatibility, or stable Codex compatibility.

Abstract: Code optimization plays a crucial role in the development of virtual machine compilers, with optimization frameworks significantly enhancing the performance of generated assembly code. However, existing virtual machine compiler outputs frequently exhibit redundant computations, inefficient loop structures, and suboptimal function implementations, which collectively impair execution efficiency. To address these shortcomings, we propose CompileRover, an advanced optimization framework specifically designed for virtual machine compilers. CompileRover employs a sophisticated three-role collaboration mechanism, comprising a referee, an advisor, and an operator, effectively overcoming performance bottlenecks by leveraging comprehensive optimization algorithms and novel methodologies, including control flow analysis, code structure transformations, and dynamic execution pattern recognition. Extensive evaluations demonstrate that CompileRover consistently surpasses state-of-the-art virtual machine compilers, achieving significant improvements in execution performance across various benchmarks. Furthermore, performance analyses validate that the introduced optimizations notably reduce execution overhead, improve dataflow consistency, and robustly enhance compiler performance, showcasing CompileRover as an effective and reliable approach to optimizing virtual machine compilers.

Abstract: Credit card fraud detection remains a challenging classification problem due to the severe imbalance between legitimate and fraudulent transactions, where conventional accuracy can provide a misleading assessment of model effectiveness. This study proposes an explainable machine learning model for credit card fraud detection by comparatively evaluating Logistic Regression, Random Forest, XGBoost, LightGBM, and CatBoost against a Dummy Classifier baseline. Model performance was assessed using balanced accuracy, precision, recall, F1-score, ROC-AUC, PR-AUC, confusion matrices, prediction errors, and computational training time. McNemar’s test was further employed to statistically assess differences in paired model predictions, while SHapley Additive exPlanations (SHAP) provided global and local interpretability of the best-performing model. Experimental results identified XGBoost as the strongest overall classifier, achieving a PR-AUC of 0.8231, ROC-AUC of 0.9752, fraud precision of 0.9259, recall of 0.7895, and F1-score of 0.8523. XGBoost additionally produced the fewest prediction errors and the shortest training time among the evaluated gradient-boosting models. McNemar’s test indicated a statistically significant difference in paired model predictions (p = 1.29 × 10⁻¹¹). SHAP analysis identified V14, V4, V12, V10, and V11 as the most influential features, while transaction-level explanations revealed how individual feature values contributed toward or against fraud predictions. The findings demonstrate that combining gradient boosting, imbalance-sensitive evaluation, statistical validation, and explainable AI can provide an accurate, computationally efficient, and interpretable approach to credit card fraud detection.

Abstract: This study addresses the growing need for vocational graduates in the construction industry to possess practical, job-relevant competencies amid rapid digital transformation. It examines the effectiveness of a BIM–XR-based virtual site inspection approach for competency-based assessment in vocational civil engineering education. Using a quasi-experimental design, students engaged in immersive inspection activities developed from BIM models and delivered through Extended Reality (XR). A weighted assessment framework evaluated six competency domains: structural inspection accuracy, procedural compliance, safety and quality awareness, decision-making, digital interaction skills, and task efficiency. Pre- and post-test analyses using descriptive and inferential statistics revealed a statistically significant improvement in overall competency scores, with a large effect size. Significant gains were particularly evident in inspection and procedural competencies, which are critical in professional practice. Additionally, the distribution of competency levels shifted, with most students achieving high competency after the intervention. The findings demonstrate that integrating BIM and XR enhances the authenticity and effectiveness of competency-based assessment. This study contributes to TVET and BIM–XR literature by providing empirical evidence that immersive virtual environments can support not only learning but also performance-based assessment aligned with real-world construction practices.

Abstract: Attribute reduction is a fundamental problem in rough set theory that seeks a minimal subset of condition attributes preserving the discernibility capability of the original decision system. Traditional heuristic methods—greedy forward selection and backward elimination—suffer from two critical deficiencies: the absence of principled attribute-ranking criteria (which leads to over- or under-selection) and the omission of statistical discriminability from the search process. To address these issues, this paper presents MFDRRS, a hybrid framework that integrates three complementary strategies: (1) Fisher discriminant ratio (FDR) evaluation of attribute discriminability based on between- and within-class scatter; (2) FDR-based filtering to select statistically significant candidate attributes; and (3) greedy forward selection guided by the rough-set dependency degree, enhanced with an FDR-driven termination criterion. FDR computation quantifies the discriminative power of individual attributes, thereby facilitating the selection of an optimally discriminative candidate subset. Since the traditional FDR is defined for continuous features, we adapt the continuous-domain FDR to nominal data by proposing a modified Fisher discriminant ratio (MFDR). The MFDR first maps nominal attributes via one-hot encoding and then computes the between-class and within-class scatter matrices. The subsequent FDR-filtering and rough-set dependency analysis guarantee the optimality of the selected attribute subset. Evaluated on synthetic and real-world UCI datasets against six baselines—classical rough-set greedy forward/backward selection, exhaustive search, MI ranking, MaxRelevance, and JMI—the proposed method achieves more compact reductions with fewer attributes, maintains competitive dependency degrees, and avoids the over-selection characteristic of classical backward elimination.

Abstract: Smart-building structural health monitoring (SHM) requires a unified digital representation capable of integrating heterogeneous sensing devices, continuous structural states, and burst-oriented post-event assessment without embedding device-specific logic throughout the software stack. This study proposes a semantic digital twin architecture in which SensorType, DeviceProfile, and site metadata form a semantic single source of truth and generate W3C Web of Things Thing Descriptions at runtime. The resulting WoT-driven contract governs field mapping, schema-on-write persistence, generic API access, state visualization, and engineering-threshold evaluation. To accommodate heterogeneous temporal behavior, event-driven seismic assessment and state-driven construction tilt monitoring are orchestrated as distinct workflows that share persistence, notification, and observability services while retaining separate timing contracts. Controlled extension experiments required no manual data-layer, backend, ingestion, or frontend modification, with a runtime source-hash difference of zero. Under a ten-building seismic-event burst, continuous write-lag p95 changed by −20 ms from a 969 ms baseline while all event jobs completed without restart or out-of-memory conditions. The ingestion path further sustained 71,040 points/s at 300 sensors with no dropped points. These results demonstrate that WoT-driven semantic interoperability and event–state workflow orchestration can provide an extensible integration foundation for smart-building SHM within a clearly defined configuration boundary.

Abstract: Auditing a firewall as a black box means learning, from accept/deny answers alone, whether a device implements the intended policy. A companion paper showed this is inherently expensive classically: a hidden rule in a header space of size N = 2^104 forces N − 1 queries, while structured tree-rule and disjoint listed-rule policies are learned with Θ(m log W) queries (m rules, field width W). We ask what changes when the auditor is quantum. If the firewall records which rule it hit (a hit counter), every quantum advantage vanishes: the counter is a projective measurement of the query register. Quantum queries do not help on structure: localising a tree-rule or disjoint listed-rule policy still needs Ω(m log W) queries, by a direct-sum theorem for the adversary bound over m independent ordered searches. On overlapping listed-rule policies, a hidden-corner family shows that re-localising one rule whose corner lies on an antichain of width n, the other rules being known, costs Ω(n) classically and Θ(√n) quantumly against O(log W) for a tree boundary: in incremental audits, overlap can be exponentially more expensive than tree structure. Only unstructured tasks keep the Grover square root (Θ(√N) hidden rule, Θ(√(N/M)) witness, Θ(√m) certificate replay). Simulations with real policies compiled into quantum oracles confirm the constants.

Abstract: Electric vehicle (EV) powertrains couple electrochemical and electromechanical subsystems whose interacting degradation and thermal behaviours are difficult to observe with conventional bench testing alone. This paper presents a modular digital twin (DT) framework for EVs that couples high-fidelity, open-source battery modelling in PyBaMM (Python Battery Mathematical Modelling), based on the Doyle–Fuller–Newman (DFN) electrochemical model, with electromechanical motor simulation in the GYM Electric Motor (GEM) library for a permanent magnet synchronous motor (PMSM). A physics-based vehicle-dynamics layer converts a commanded speed profile into power demand and coordinates energy flow between the two subsystems in real time. The framework was exercised over standardized highway, urban, and combined drive cycles and evaluated in terms of state of charge (SoC), state of health (SoH), remaining useful life (RUL), torque, efficiency, and thermal behaviours. An interactive web-based dashboard exposes these metrics for simulation control and visual diagnostics. Correlation analysis of the resulting time-series confirms physically consistent relationships among key variables, notably a near-unity correlation between vehicle speed and motor speed and a strong thermal coupling (r = 0.91) between battery and motor temperature, which supports the framework’s validity as a basis for predictive maintenance and control-strategy research. The results indicate that coupling open-source, physics-based battery and motor solvers within a single interactive DT is a practical and extensible route to EV powertrain diagnostics that does not depend on proprietary simulation suites.

Abstract: Security operations centres (SOCs) face two simultaneous mandates: keep the network traffic they analyse private, and make their AI-based detectors explainable. These goals are usually assumed to be in tension, because in image and language models differential privacy (DP) is reported to degrade post-hoc explanations. We present the first systematic audit of explanation faithfulness under formally-accounted (ε,δ)-DP in federated, tabular deep intrusion detection (IDS), and we add a previously-unstudied axis: the interaction between Byzantine-robust aggregation and explanation quality. Across two datasets (ToN_IoT, CICIDS2017), two architectures (an MLP and a compact tabular transformer), three attribution methods (Integrated Gradients, GradientSHAP, attention rollout), privacy budgets ε ∈ {8, 4, 1} (ε ∈ {4, 1} for the transformer), and heterogeneity levels, and using paired Wilcoxon tests with Cliff’s δ effect sizes, we find — contrary to our own anchor hypothesis and to the image/NLP literature — that client-side DP-SGD does not degrade explanation faithfulness. Measured by comprehensiveness and the AOPC MoRF−LeRF separation it significantly increases self-faithfulness (large δ, all seeds), an effect that attenuates at the strongest budget and survives a confidence-ceiling (f0) normalization. A pre-registered capacity-matched control attributes this faithfulness gain largely to DP’s utility cost — non-DP models throttled to the same accuracy by three independent mechanisms reproduce it — leaving only a small, clipping-driven residual in the centralized setting. DP’s real cost falls on utility and, most sharply, on rare-attack detection (mitm F1 0.65→0.05). Per-class utility and faithfulness decouple: well-supported classes hold their F1 while their attribution gap widens. A strong model-poisoning attack collapses FedAvg in both utility (0.83→0.20) and faithfulness (gap 0.58→0.21), whereas median, trimmed-mean and multi-Krum preserve both — though median/Krum cost some faithfulness when no attack is present, making trimmed-mean the best all-round trade-off. Among attribution methods, attention rollout is the most fragile. Practically: for explainable federated IDS, gradient/Shapley attributions remain trustworthy under DP, and trimmed-mean aggregation buys robustness at little explanation cost.

Abstract: Road traffic crashes remain a major public safety concern in Ecuador, particularly on the General Rumiñahui Highway, a high-demand corridor characterized by complex geometry, variable traffic density, and operational constraints. This study proposes a three-layer hybrid framework to support real-time accident risk prediction and prevention in this corridor. The architecture integrates Internet of Vehicles (IoV) data acquisition through roadside and vehicular sensors, IP cameras, and environmental sensors; Multi-access Edge Computing (MEC) for local processing, YOLOv8-based visual perception, and V2V, V2I, and V2P communications; and cloud-based predictive analytics using a CNN–LSTM model. The framework is aligned with ISO 37106:2021 and ISO 37122:2019 to promote interoperability, smart-city operational integration, and mobility-related performance indicators. Its ex ante validation relies on comparative evidence from studies of equivalent architectures and on the consistency of available institutional data sources. Reported evidence supports the technical plausibility of visual detection, spatiotemporal risk modeling, and low-latency MEC-enabled V2X communications; however, these results do not constitute empirical performance measurements for the General Rumiñahui Highway. The study concludes that the proposed framework is a technically feasible basis for a future instrumented pilot, subject to institutional integration, sufficient sensing infrastructure, connectivity, and field validation.