Computer Science and Mathematics

Abstract: Autonomous AI agents now hold funds, delegate authority to other agents, and transact at machine speed; the governance apparatus meant to constrain them—policies, audits, compliance—remains documentation-based and human-latency. This mismatch cannot be closed by better monitoring or filtering: compliance must become a runtime, compositional, proof-carrying property of computation itself. We call the resulting discipline computational jurisprudence. This article surveys the four literatures the discipline must synthesize: object-capability security; verifiable, proof-carrying, and zero-knowledge computation; policy-as-code and computational law; and agentic AI with its emerging payment protocols. Each supplies a mature mechanism the others lack; none supplies a complete normative substrate. The synthesis is organized in three pillars: (i) a delegation calculus under which authority can only attenuate as it propagates between agents; (ii) runtime compliance proofs, a three-tier evidence regime (attested, optimistic, and zero-knowledge); and (iii) sealed delegation chains with graduated attribution, which reconcile the privacy of capability-based authority with the accountability that adjudication requires. A case study on agentic payment protocols grounds the architecture and reports first measurements: capability verification versus a centralized policy decision point, end-to-end enforcement on the x402 payment path, and accumulator-based revocation. Seven open problems define the research agenda.

Abstract: This article presents the development of an optimized cross-platform post-quantum cryptographic (PQC) library that enhances the cryptographic security of data, networks and IT infrastructures in real-world organizational environments. Classical cryptographic standards such as RSA and ECC face existential threats from the advent of quantum computing, necessitating an urgent transition to quantum-resistant cryptographic algorithms. By optimizing execution speed through AVX2 vectorization via compiler intrinsics and ensuring portability across different operating environments, this work demonstrates the feasibility of deploying high-performance PQC in today's operational systems. The approach bridges PQC research to practical systems by integrating ML-DSA and ML-KEM implementations into a single compiled library, while adapting PQC implementations for cross-platform portability. Benchmarking results demonstrate up to a fourfold improvement in execution speed over reference implementations. This work strengthens an organization's security posture in the face of quantum computing threats and enables long-term cryptographic resiliency across diverse platform environments.

Abstract: While Federated Learning offers a decentralized approach to model training, ensuring the integrity of the information from each IoT client remains a challenge. This work delves into the dynamics of Multi-Stage Federated Learning, its susceptibility to information integrity attacks, and how to defend against such threats. A comprehensive understanding of data uncertainty and the challenges of poisoning attacks is discussed, laying a solid groundwork for the proposed defense mechanisms. At its core, this paper introduces a novel Multi-Stage Federated Learning model that segments the Federated Learning process into distinct phases with a novel approach of inertial momentum aware Inexact Augmented Lagrange Multiplier Robust PCA with constant momentum factor and unaltered norm of the traditional one, each tailored to optimize for both efficiency and security. This robust framework is then tested against data injection based poisoning attacks, using sparse noise, and demonstrates the effectiveness of the proposed recovery techniques like Robust PCA. Performance results highlight the resilience and efficacy of the introduced model with novel reconstruction algorithm, emphasizing the importance of this approach in real-world IoT settings. Data analysis, model summaries, and impacts of adversarial attacks further reinforce the findings, which are evaluated using rigorous statistical metrics and machine learning algorithms. The paper concludes by acknowledging its efficacy in detection and recovery from data poisoning attacks, improving robustness and data reconstruction in IoT environments while highlighting opportunities for further security enhancements.

Abstract: Proactive cyber defense requires defenders to anticipate attacker behavior and respond before breaches occur. However, existing game-theoretic models rarely integrate attack behavior simulation with temporal-aware payoff structures. This paper proposes a Stackelberg game framework for proactive cyber defense that combines Monte Carlo-based attack behavior simulation with a multi-stage, time-discounted payoff function. The defender, modeled as the Stackelberg leader, commits to an optimal mixed strategy by leveraging simulation-derived attack path predictions across five Cyber Kill Chain (CKC) stages over a 5-action defense space (|A_D| = 5) against five attacker actions (|A_A| = 5). A simulation confidence factor S quantifies prediction accuracy and is embedded directly into the defender payoff function, creating incentive-compatible alignment between simulation quality and strategic gain. Bayesian belief updates—augmented with simulation evidence—enable real-time attacker type inference under incomplete information. A fictitious play algorithm with provable convergence computes the Strong Stackelberg Equilibrium (SSE). Experimental results on a 10-node network demonstrate that the proposed framework reduces response latency by 22.9% and achieves a 4.9% payoff gain compared to Bayesian Nash Equilibrium (BNE) baselines, with convergence at 199 iterations versus 354 for BNE—a 43.8% reduction.

Abstract: Cross-organization cyber defense must reconcile collaborative learning with privacy and adversarial robustness — yet standard federated learning ships full gradient tensors, leaking sensitive posture and inviting Byzantine manipulation. We present FedMARL-LTI, a federated multi-agent reinforcement learning framework whose architecture answers both pressures with a single decision: organizations share neither raw data nor model weights, only differentially-private 768-dimensional semantic threat embeddings. The contribution is fourfold. (1) Semantic Abstraction (SA) channel: per organization, each round, the local gradient is summarized by an LLM, projected to a 768-dim embedding, L2-clipped, and Gaussian-noised before any numeric quantity leaves the host. The bottleneck reduces the per-element noise scale from O(√(d_model )) to O(√m) with m=768≪d_model≈3×10⁵. (2) Formal privacy analysis: the SA+DP cascade satisfies (ε,δ)-DP and bounds per-round mutual-information leakage by min{T_toklog₂V, m⁄2 log₂ (1+C²/(mσ²))} (Theorem 1), with Rényi composition over T federation rounds (Theorem 2). (3) Byzantine-resilient ClippedClustering aggregator combining L2 clipping with cosine-similarity clustering. (4) Hierarchical MARL policy with threat-profile-aware LLM-IRR reward shaping, wired end-to-end and disclosed honestly (the LLM call is currently stubbed with a deterministic projection for reproducibility). We evaluate on CybORG CAGE-4 with n=5 organizations, 30 federation rounds × 5 episodes × 100 steps per round. The SA channel adds statistical-zero utility cost vs. no-privacy baseline: SA-only Δreward = -0.66 (t=+0.31, NS), dual SA + Weight-DP Δreward = +1.90 (t=-0.71, NS), all N=5 seeds, all |t|< 1.3. A controlled signal/noise probe confirms a 19.58× improvement of SA over Weight-DP at fixed DP budget — matching the predicted √(d/m)≈19.8. Under Byzantine sign_flip at 30% (N=15), ClippedClustering is directionally strongest (F₁=0.025 vs FedAvg 0.020, Krum 0.016) but the edge is not statistically significant (CC vs Krum t=+1.59, p=0.15, d=+0.58; the earlier N=5 “3.4×” gap was small-sample optimism, §5.2); its decisive Byzantine win is the harsher random_noise attack, where FedAvg diverges to NaN and Krum collapses to 0.002 while ClippedClustering survives at 0.020 (§5.7, Cohen’s d=+3.77). The cooperative-PPO family (MAPPO, IPPO) outperforms value/actor-critic (QMIX, MADDPG) by ≈20 reward units, p< 0.001. All host-level F1 values stay below 0.05 at the 15K-step training horizon used here; the relative claims of the paper (privacy zero-cost, ClippedClustering’s decisive Byzantine win on the harshest attacks per §5.7, cooperative-PPO dominance) are unaffected by this scope. A 200K-step long-horizon replication (§6.3 L1) lifts F₁ above the 15K plateau (to ≈0.044, N=5) — confirming that horizon, not the privacy/Byzantine machinery, gates absolute accuracy — but a finer 60-checkpoint run shows the climb is volatile and non-monotonic and does not reach deployment-grade, an honest stability-not-compute limitation. We release all 141 raw run JSON outputs (Phases 1–3, the L4 backend comparison, and the algorithm/aggregator baselines), the figures, and analysis scripts for replication.

Abstract: Translating audit requirements into computable specifications for trusted digital infrastructure remains difficult. A major obstacle is the semantic disconnect between audit professionals and system architects during fintech system development. An additional challenge is embedding executable controls early with an appropriate degree of integration, rather than treating compliance mechanisms as an afterthought during later coding, integration and delivery across complex organizational and technical settings today. Drawing from process walkthroughs and software control mapping practices in internal auditing and systems examination, this study creates an audit evidence type and technical counterpart classification model, a mapping conversion matrix from audit assertions to system functions, and a verifiability-oriented control effectiveness verification algorithm. The methodology was illustrated through a practical digital asset management platform case. These results show that using the method in an architectural design phase creates a closed-loop constraint on controls, functions and evidence, decreasing interface changes, data fixups and test burdens arising when remedial actions are taken after development. It added a practical route to computer-engineering implementation for embedded compliance architecture and audit-ready system design within current software engineering practice.

Abstract: The Model Context Protocol (MCP) enables LLM agents to interact with external tools, resources, file systems, and platform features. Although this improves interoperability in agent-based systems, improperly configured MCP servers may expose serious security risks, including unauthenticated tool invocation, excessive file access, arbitrary command execution, prompt injection, token replay, response leakage, and denial-of-service behavior. The current work presents a controlled security assessment of an insecure MCP server and evaluates a lightweight FastAPI-based security wrapper that hardens the server without modifying its original implementation. The wrapper mechanism enforces JWT authentication, role-based tool allow listing, input validation, prompt injection filtering, threat scoring, response inspection, rate limiting, and audit logging. The evaluation used 82 controlled scenarios across nine test suites mapped to STRIDE threat categories, including 69 attack cases, 10 rate-limit/DoS cases, 3 false-positive checks, and 3 legitimate-use validation cases. Compared with the insecure baseline, the wrapper reduced vulnerability exposure by 96%, blocked all tested attacks in several categories, and achieved 100% scenario-level mitigation for rate-limit exhaustion. The wrapper mechanism introduced low overhead, with an average latency increase of approximately 3 ms, corresponding to 1.1%, and a memory overhead of 4.3 MB, corresponding to 8.4%. These results show that external wrapper-based hardening can provide practical deployment-level protection for MCP servers, while remaining a complementary defense layer rather than a complete security solution.

Abstract: The Post Office Horizon scandal represents one of the most severe miscarriages of justice in modern British legal history, rooted in the deployment of a defective IT system and the institutional suppression of evidence that exposed its unreliability. This paper provides a systematic analysis of the scandal through the combined lenses of cyber security, digital forensics, and IT governance, drawing directly on the Post Office Horizon Public Inquiry dataset—including witness testimony, technical documentation, audit records, and internal communications spanning more than two decades. While many of the technical and procedural failures discussed have been documented in prior scholarship, the paper’s principal contribution is the systematic mapping of those failures against recognised governance frameworks, and the Legally-Accountable Digital Systems (LADS) proposal this mapping motivates. We identify and analyse five interconnected failure categories: software defects and poor system design; deficient patch governance; inadequate audit logging and compromised evidence integrity; investigative failures and prosecutorial conflict of interest; and a systemic absence of technical expertise and independent oversight. Each category is mapped against nine governance framework documents, including ISO/IEC 27001, NIST SP 800-53 Rev. 5, NIST SP 800-218 (SSDF), COBIT 2019, ISO/IEC 27035, NIST SP 800-61, and ISO/IEC 27036, with ISO/IEC 27037 applied additionally to evidential handling failures and NIST SP 800-92 to log management failures. A counterfactual analysis indicates that compliance with these frameworks could plausibly have detected, exposed, or substantially reduced most of the documented failures, though this claim is necessarily inferential and conditional on good-faith implementation. The scandal was therefore not caused primarily by the absence of adequate frameworks, but by their wholesale non-application and by institutional incentives, examined later in the paper, that can undermine even fully compliant controls. However, the analysis also surfaces a governance gap that no existing framework addresses: the institutional failure mode in which the organisation responsible for system integrity holds active incentives to suppress evidence of failure rather than remediate it. To address this gap, we propose the concept of Legally-Accountable Digital Systems (LADS) — a governance category for systems whose outputs are used as evidence in legal proceedings — and outline three supplementary pillars: technical independence, institutional independence, and forensic admissibility governance. We draw a parallel with the Sarbanes-Oxley Act of 2002, arguing that the Horizon Inquiry dataset provides an equivalent empirical foundation for statutory reform of digital evidence governance. Finally, we outline the substantial research opportunities the Inquiry dataset presents across IT systems analysis, cyber security, forensic accounting, social network analysis, and legal informatics — a resource comparable in significance to the Enron materials that shaped a generation of corporate governance reform.

Abstract: The rapid adoption of Quick Response (QR) codes across digital services, electronic payments, public infrastructure, and everyday interactions has significantly expanded the attack surface for QR-code-based phishing attacks, commonly referred to as quishing. This study examines quishing from a sociotechnical perspective, analyzing how technological mechanisms, contextual trust environments, and user behavior interact to facilitate successful attacks. A qualitative and analytical methodology was used, combining a structured review of cybersecurity literature with the conceptual analysis of documented quishing attack scenarios. The study identifies the principal stages involved in quishing attacks and examines the sociotechnical conditions that increase user susceptibility, particularly the presence of seemingly trustworthy environments, routine low-risk user actions, and the limited visibility of verification mechanisms prior to QR-code interaction. Based on these findings, the study develops a sociotechnical model of quishing attacks, a user security decision flow model, and an integrated conceptual framework linking attack dynamics, user decision processes, and protective strategies. Additionally, a Decalogue of Essential Protection Against Quishing is proposed, providing practical user-centered recommendations aimed at reducing vulnerability to malicious QR codes. The findings suggest that mitigating emerging QR-code-based phishing threats requires not only technological countermeasures but also behavioral awareness, verification practices, and user-centered cybersecurity strategies capable of addressing both digital and physical interaction contexts.

Abstract: Local governments increasingly rely on digital infrastructure to deliver public services, yet research on their financial response to cyberattacks remains limited. This study introduces the Cyber-Incident Budget Response (CIBR), capturing IT and cybersecurity budget adjustments following an attack, and examines whether such responses reflect strategic planning or reactive decision-making grounded in Threat Rigidity Theory. Budget data spanning fiscal years 2015–2022 were collected for 57 U.S. local governments victimized by ransomware in 2019 and 57 matched non-victim governments. A robust regression (MM-estimation) tested the relationship between attack occurrence and budgetary response, controlling for population, income, and racial diversity, with income as a moderator. Results show a significant positive relationship between cyberattack occurrence and budgetary response, confirming the existence of a CIBR. Racial diversity was independently associated with stronger responses, and the attack-income interaction was significant and negative, indicating the budgetary impact diminishes as income rises. Findings suggest local government cybersecurity spending is often reactive rather than strategically planned, with recovery costs falling disproportionately on lower-income jurisdictions. The study offers a novel framework for examining post-incident fiscal behavior and implications for equitable cybersecurity policy.

Abstract: Acoustic side-channel attacks (ASCAs) exploit unintended sound emitted by keyboards to infer typed input, but existing methods generally assume manually-labelled training data and controlled environments, limiting their applicability to realistic scenarios such as online lectures. We develop a pipeline that automatically labels keystroke-sound samples captured from online coding tutorials: video frames are processed with optical character recognition (OCR) to extract the ground-truth character sequence, audio is segmented into clips centred on detected click events, and the two streams are aligned. A convolutional neural network (CNN) is trained on mel-spectrogram features, with transfer learning used to adapt the pretrained model to a target user with minimal samples. Our dataset contains 50 unique keys from standard QWERTZ keyboards recorded during real programming lectures. On a held-out test set the CNN achieves 98.1 % top-1, 99.4 % top-2 and 100 % top-3 accuracy. Transfer learning retains strong performance with as few as 13 samples per key. Pairing OCR-derived ground truth with acoustic CNN classification removes the labelling bottleneck that has limited previous ASCAs, and the transfer-learning stage makes the attack viable with minimal per-victim data. All code, trained models, and labelled datasets are released to support reproducible research.

Abstract: Few-shot domain adaptation (FSDA) has become a key approach for cross-domain intrusion detection, enabling models to leverage limited labeled target data under distribution shift. While a wide range of adaptation methods have been proposed, their effectiveness often varies significantly across different transfer scenarios, leading to inconsistent performance and limited interpretability. In this work, it is argued that such variability stems from an overlooked factor: transfer difficulty. This study proposes a transfer-difficulty-aware perspective on FSDA and shows that adaptation behavior is fundamentally dependent on cross-domain compatibility rather than solely on intrinsic domain structure. To this end, a distinction is made between intra-domain separability, which characterizes the internal structure of each domain, and transfer difficulty, which captures how well source-derived representations generalize to the target domain. A set of asymmetric transferability metrics is introduced to quantify this phenomenon and accompanied by a systematic analysis across multiple transfer directions. The results reveal that a domain with strong internal separability does not necessarily yield easy transfer, highlighting that intra-domain structure alone is insufficient to explain cross-domain performance. Furthermore, it is shown that different adaptation strategies exhibit distinct behaviors depending on transfer difficulty: target-only few-shot learning is effective in low-difficulty settings, whereas alignment-based approaches become essential in high-difficulty scenarios. These findings explain the inconsistent performance of existing methods and suggest that domain adaptation should be treated as a transfer-dependent problem rather than a uniform strategy. Ultimately, this work offers both theoretical insights and practical guidance for designing robust cross-domain intrusion detection systems.

Abstract: In multi-asset UAV safety monitoring scenarios, dynamic threat assessment is challenged by delayed recognition of protected-asset shifts and unstable threat rankings under noisy observations. To address these issues, this paper proposes an approach-directionality-based dynamic UAV threat assessment method that integrates historical trend correction and adaptive exponential moving average (EMA) smoothing. The proposed method takes each UAV–protected asset pair as the basic assessment unit and constructs a threat matrix for multiple UAVs and multiple protected assets. Based on fundamental threat factors, including UAV type, closing speed, altitude, and distance, an approach directionality factor is defined using the geometric relationship between the UAV velocity direction and the bearing direction of each protected asset. To capture shifts in the potentially affected protected asset, the historical directionality sequence is linearly fitted, and the instantaneous directionality is dynamically enhanced or suppressed according to the estimated trend. An adaptive EMA mechanism is then introduced to improve temporal stability under noisy conditions. Simulation results from turning-approach scenarios, Monte Carlo experiments, and noise robustness tests show that the proposed method identifies protected-asset shifts earlier while maintaining high protected-asset identification accuracy and ranking stability. The framework provides interpretable support for threat ranking, situational awareness, and safety-oriented decision-making in non-offensive UAV safety monitoring applications.

Abstract: As enterprises continuously rely on data to effectively drive and power business processes in a digital economy, privacy concerns have emerged as a major source of deep concern among consumers and privacy advocates in a digital society. Several scholars have shared their opinions and perspectives in related articles on this issue since privacy is a fundamental and constitutional right in many countries that must be protected at all times. While several practitioners and industry experts have proffered various privacy-preserving measures to help empower users to make informed decisions that relate to the use of their personal information, others have proposed various privacy preserving measures and mechanisms to help protect the use of personal information in a digital society to create the necessary confidence and trust amongst members of the public. With the current advances in artificial intelligence, social media platforms, and automation, the issue has emerged as a major source of concern among policymakers, privacy advocates, and industry experts. The purpose of the e-Delphi study was to gain consensus from the opinions of industry experts on best practice measures and various effective privacy-preserving measures to help enhance users’ privacy and allay privacy concerns in a digital society. The study adopted the Restricted Access/Limited Control (RALC) theory of privacy to provide a theoretical framework for the research study. The study included three rounds of questioning using the Delphi method. The findings from the study revealed that effective privacy-preserving measures, such as Data minimization, Privacy-By-Design, Privacy Labels and icons, Data Ownership and control, Third-party App Permission, Mandatory Data/Privacy Breach Notice, Frequent Policy Updates, End-to-End Encryption, User-Friendly Privacy Control features, and Informed Consent, provided an effective way to allay the fears of consumers in a digital age.

Abstract: The use of industrial control systems (ICS) in the nuclear sector is widespread and is already considered one of the critical assets to evaluate in defining possible attack surfaces for malicious actors linked to crime or hostile nations. The risk of insider threats, both internal and related to equipment suppliers and the supply chain, is always high and must be monitored in dedicated manner with specific protocols, penetration tests and stress analysis. The typical approach to cybersecurity in IT (information technology) systems is not suitable for managing Operational Technology (OT) systems security. The advent of artificial intelligence (AI) is certainly an additional tool for enhancing defense, but it can also be a tool for possible attacks and a risk for the nuclear system in its intimate control of critical process. In this paper, we intend to review and compare various relevant standards that offer deep and comparable rules on ICS and general issues pertaining to the use of AI in most systems contextualized in nuclear domains, what institutions and international organizations are doing to exploit its advantages, and highlight the possible risks, referred to standards and guidelines.

Abstract: Autonomous sensor networks and edge-based intelligent systems require lightweight cryptographic mechanisms capable of ensuring integrity, auditability, and replay resistance under constrained computational resources. This paper proposes a Quantum-Entropy-Assisted Hybrid Commitment (HQC) framework that integrates QRNG-derived entropy into a classical hash-based commitment architecture designed for embedded sensing environments. Experimental evaluation on a Raspberry Pi 4 achieved mean commit latency of 5.1 μs, verify latency of 2.2 μs, and Merkle root construction of 1.640 ms for 1024 leaves. Security validation over 10,000 trials yielded zero forgeries, complete replay rejection, and no statistical distinguishability (p=0.7947).The results show that a quantum-entropy-assisted commitment can be practically implemented on resource-constrained embedded systems without the use of infrastructure for quantum communication or quantum computation.

Abstract: Machine learning intrusion detection systems are routinely evaluated by randomly splitting a single dataset into training and test partitions and reporting accuracy or F1-score. This practice silently assumes that training and test traffic share one data-generating process, an assumption that fails whenever attacks, networks, or capture conditions change. This paper presents a controlled empirical study of concept drift in two widely used flow-based benchmarks, CSE-CIC-IDS2018 and CIC-DDoS2019, and shows that the evaluation protocol itself is the dominant variable governing reported performance. Under a random split, Random Forest, LightGBM, and a multilayer perceptron all reach macro-F1 ≈ 0.79–0.82 on CSE-CIC-IDS2018; under a class-blind chronological split the same models collapse to macro-F1 ≈ 0.02, a gap of 0.78 (95% bootstrap CI [0.75, 0.80], empirical p < 0.001). A label-shuffle control confirms the pipeline is leakage-free (macro-F1 0.065, at the 1/15 chance level). We further measure cross-dataset transfer from CSE-CIC-IDS2018 to two distinct CIC-DDoS2019 capture regimes: zero-shot transfer collapses in both (Matthews correlation coefficient −0.043 and −0.078, worse than chance). A small time-ordered target calibration buffer then repairs confidence reliably in both regimes (expected calibration error falling to ≤ 0.11 at the largest (10%) buffer) but restores discrimination unevenly, weakly in one regime (MCC to ≈ 0.26) and substantially in the other (MCC to ≈ 0.71), with the calibration metrics giving no indication of which outcome occurred. Repairing confidence therefore does not guarantee a usable model. A distributional analysis attributes the collapse to pervasive feature drift: 36 of 45 common-core features exceed the severe-drift threshold of PSI ≥ 0.25, with Kolmogorov–Smirnov statistics up to 0.79. We conclude that a credible intrusion-detection evaluation must report time-aware splits, cross-regime transfer, and calibration quality, not single-split accuracy.

Abstract: Intrusion detection in microgrid systems is a cyber-physical task that requires correlating different data from networks, hosts, and endpoints to create actionable evidence. Existing approaches largely treat intrusion detection as a classification problem and provide explanations at the sample or feature level. However, these explanations lack physical interpretability and fail to reveal cross-modal interactions underlying system decisions. As a result, operators cannot reliably trace detected anomalies to the physical layer, limiting the ability to diagnose root causes. This leads to incorrect or delayed responses and potentially compromises the safety of microgrid operations. This work proposes a physical and data-link layer explainable intrusion detection framework via cross-modal evidence reasoning. This framework reformulates intrusion detection as an operation Q\&A task over structured multi-modal evidence, including network flows, Software-Defined Networking (SDN) states, system calls, and power measurements. By designing an evidence-based explanation mechanism, sample importance is aligned with structured evidence and aggregated into physical modalities to construct evidence representations. These representations are further transformed into structured features to build joint decision models, enabling the extraction of decision paths and their conversion into interpretable reasoning processes grounded in physical evidence. The proposed framework is evaluated on realistic cyber–physical microgrid datasets. It provides consistent and physically meaningful explanations, revealing distinct cross-modal evidence patterns across different cyber attacks. This work advances intrusion detection from samples to physical-layer reasoning, enabling trustworthy security analysis in microgrid systems.

Abstract: Recent cyber incidents have become increasingly sophisticated through 'Living-off-the-Land (LotL)' techniques that exploit legitimate behavior and multi-stage attacks. This demands advanced reasoning capabilities to discern attack contexts within fragmented, large-scale logs. However, closed network environments with physical network separation (air-gapped), such as national critical infrastructure, restrict the use of high-performance cloud LLMs, limiting the adoption of cutting-edge AI-based analysis technologies. This research proposes a Local LLM-based intrusion analysis framework that can operate independently within closed networks to overcome these constraints. The proposed framework combines (i) an Offline Knowledge Distillation technique that transfers the analytical reasoning process of external high-performance models to the Local LLM after security review, and (ii) an AI agent orchestration structure that controls the analysis procedure step-by-step and suppresses hallucinations. Experiments and validation using the public dataset (Atomic Red Team) demonstrate that the proposed model achieves significantly higher detection accuracy (88.4%) and MITRE ATT&CK mapping performance (0.91 F1-Score) compared to existing general-purpose Local LLMs. Furthermore, it suppressed hallucination rates to 6.2% through an automated verification mechanism and significantly improved analysis efficiency by refining large-scale logs to focus on core events. This study quantitatively demonstrates that AI-based intrusion incident analysis automation is achievable using a single GPU server even under the resource constraints of closed networks, presenting a practical solution for intelligent security monitoring.

Abstract: Operational technology (OT) and information technology (IT) have become increasingly integrated, expanding the attack surface of OT systems. Power from shore has also become more widespread for offshore critical infrastructure, and has introduced new dependencies and the potential for a single point of failure. In addition, the cyber threat landscape is escalating, with state-sponsored actors demonstrating the capabilities and willingness to target industrial systems. Threat actors have been seen using living off the land techniques, such as with the Industroyer malware, which utilized legitimate but malicious IEC 104 commands. To evaluate these vulnerabilities, this study applies a Design Science Research approach to map a generalized substation and develop a Software in the Loop simulator. The simulator was used to test specific attack vectors against substation automation systems. The results confirm that an adversary with local network access can successfully inject valid IEC 61850 MMS commands to trigger unauthorized circuit breaker operations. Furthermore, the results show that it is possible to use a simulated substation as a tool when developing ICS malware. These findings demonstrate that common operational technology protocols lack fundamental security by design, meaning the technical barrier to execute a disruptive attack is low once network access is achieved. Protecting these critical environments requires a robust defense-in-depth strategy that accounts for supply chain risks and enforces strict network segmentation.