This paper evaluates a combined methodological approach for structured IoT security assessment, integrating the OWASP IoT Security Testing Guide (ISTG), which defines what to test across hardware, firmware, network services, and communication interfaces, with the five-phase penetration testing model of the German Federal Office for Information Security (BSI), which defines how to proceed. The combination is evaluated through a real-device case study of a commercially available low-cost smart home camera. The assessment produced a verified end-to-end attack chain from physical access to full system compromise and recovery of private video data, demonstrating that the combined methodology can drive a non-trivial real-device assessment to completion. More importantly, the case study exposed integration gaps the combination does not resolve. We describe the adaptations that were necessary to bridge these gaps, present a per-test-case documentation scheme and a consolidated gap-adaptation overview as reusable artifacts, and formulate concrete recommendations for future revisions of IoT-specific test catalogues.
Operators of critical infrastructure increasingly rely on heterogeneous catalogs of cybersecurity services (probes, intrusion detection systems, monitoring tools), yet selecting and reasoning about these services remains a manual, expert-driven task. Digital Twins (DTs) have emerged as virtual replicas of physical systems for monitoring and simulation, and platforms have been built to manage the catalogs of cybersecurity services that can be deployed into such twins; however, the interaction layer of these platforms typically does not provide a reasoning interface above the catalog. To address this, we propose a smart digital twin augmented with cooperating LLM-based agents that perform knowledge-grounded reasoning, automated deployment, and incident response. As a first step, we present a Boss Agent embedded in an existing DT management platform. This conversational assistant uses Retrieval-Augmented Generation (RAG) to ground responses in the platform’s service catalog. To preserve data government in regulated sectors, the agent runs entirely on local infrastructure using open-source models. We detail the system architecture and a lightweight, embedding-margin-based intent classifier that eliminates retrieval overhead during conversational turns. Additionally, we provide an empirical evaluation on commodity hardware that measures factual accuracy, hallucination rate, and latency against a no-retrieval baseline. We conclude with limitations and a research agenda to extend the agent’s grounding to live DT states and introduce additional specialized agents.
The NIS 2 Directive and GDPR impose on regulated EU organisations a growing policy compliance gap: the obligation to produce context-specific cybersecurity policy artefacts has outpaced internal capacity to create them. Large language models offer scalable assistance, but effective policy generation requires routing sensitive organisational data through an inference pipeline. Under GDPR Arts. 28, 32, and 44–49, this creates a structural tension: the data required to generate a useful policy is precisely the data that cannot be processed without adequate legal safeguards. We present a controlled empirical benchmark comparing six AI configurations spanning the design space from unaugmented frontier chat to privacy-by-design private RAG. Sixty policy artefacts were generated across two regulated contexts—a fictitious Italian public hospital and a fictitious NATO detachment, constructed as representative synthetic profiles with no connection to real entities—and five policy types, then evaluated by an LLM-as-judge pipeline using a seven-criterion rubric derived from the Anthropic agent evaluation framework and extended with three cybersecurity compliance dimensions. The best frontier configuration (gpt-5.5 + web search) achieves a mean score of 4.60/5; the private RAG configuration (Kimi K2.5 on EU-resident infrastructure under a DPA) scores 4.15/5—a 9
The rapid adoption of open-hardware RISC-V System-on-Chip (SoC) designs in security-sensitive embedded and edge scenarios is raising the bar on observability and assurance requirements for such platforms. In fact, unforeseen runtime misbehaviors due to incompatible third-party IPs, hardware trojans in peripherals, and compromised drivers can cause misbehaviors that core-level monitors, such as RISC-V Formal Interface (RVFI), cannot see. We introduce the Peripheral Formal Interface (PFI), a passive AXI-aware observation interface that extends the RVFI philosophy from the core to the peripherals. The PFI emits a structured event stream that off-chip auditors can consume without touching the functional path. On top of it, we define an auditing architecture that merges the peripheral-bus stream and the RVFI retirement stream into a single event log, and a process mining pipeline that turns this log into a Contract: a behavioral model with allowlist, count, and timing constraints. We validate the framework on Simply-V, an open-source reconfigurable soft-SoC, auditing representative workloads to localize injected firmware defects event by event.
As maritime trade becomes increasingly digitised, the effectiveness of Carbon Border Adjustment Mechanism (CBAM) enforcement depends on the integrity of digital declarations that remain only weakly linked to the physical cargo they describe. Existing Digital Twin (DT) systems in logistics primarily focus on operational visibility and efficiency, while offering limited mechanisms for adversarial verification, provenance assurance, or compliance-oriented anomaly detection. This creates significant risks in CBAM-regulated trade environments, including manifest manipulation, route spoofing, origin fraud, and falsified emissions declarations. This paper proposes a Multimodal DT Framework for CBAM-oriented maritime trade integrity. The framework integrates customs profiling, multimodal sensor fusion, and DT synchronisation to support continuous cross-verification between declared shipment data and independently observable physical evidence. The proposed architecture combines visual inspection, weight-based verification, and vessel-tracking intelligence within a provenance-aware enforcement layer capable of detecting inconsistencies between declared and observed cargo states. To demonstrate the feasibility of the approach, a synthetic proof-of-concept simulation was developed using density-based anomaly detection aligned with HS-code cargo profiles. The results illustrate how discrepancy detection mechanisms can identify one class of manifest divergence under controlled conditions. The framework further introduces a trust-oriented DT model incorporating multimodal validation, integrity risk scoring, and closed-loop synchronisation between sensor events and customs verification systems. The study argues that future DT infrastructures for maritime logistics should evolve beyond passive visibility platforms toward enforcement-oriented compliance systems capable of supporting secure, transparent, and low-carbon maritime trade within CBAM-regulated supply chains.
Accurate prediction of daily electric vehicle (EV) energy demand supports charging planning across home users, commercial fleets, depots, and charging service providers. We study this problem as a supervised regression task under centralized and Federated Learning (FL) settings, using heterogeneous EV data sources that include two public datasets and a private commercial dataset from a postal service. Raw EV records are transformed into vehicle-day samples through a unified preprocessing pipeline based on a common schema of vehicle identifier, date, time, state of charge (SOC), and odometer readings. Daily travelled distance and energy demand are derived from these records, and the prediction target is estimated from discharge segments between charging events. The input features also combine calendar, historical demand, and weather descriptors. We evaluate local, centralized, cross-device, and cross-silo FL across three deep learning architectures, and implement a buffered asynchronous FL orchestration pipeline on a Raspberry Pi testbed to assess practical deployment. The results show that distributed EV energy demand models can approach centralized performance while preserving data locality, and that buffered asynchronous FL is practical for heterogeneous EV deployments.
Memory sanitization is a key technique for detecting memory corruption bugs during fuzzing. While tools such as AddressSanitizer are widely used in OS-hosted environments, they cannot be applied to bare-metal firmware due to the absence of operating system abstractions. We propose a sanitization architecture for bare-metal firmware fuzzing that combines static analysis of the firmware binary with runtime memory access observation inside an emulator. Built on top of the Hoedur firmware fuzzer, our approach tracks accesses to global, stack, and heap memory regions without requiring source-code instrumentation. Evaluation on 22 real-world firmware targets demonstrates that the architecture correctly reconstructs program flow and intercepts memory accesses that remain invisible to conventional crash- and hang-based fuzzing. The program flow reconstruction correctly handles diverse control-flow patterns including tail calls, loop functions, and multi-threaded execution across all targets. We trace the remaining inaccuracies to incomplete pointer resolution and unhandled bulk memory operations in the static analysis, both of which represent solvable engineering challenges. We systematically analyze the sources of imprecision encountered and identify concrete directions for achieving practical detection accuracy in future work.
Distributed Edge-Cloud systems offer scalable, flexible, and modular infrastructure but introduce significant security challenges due to heterogeneity, decentralized data processing, expanded attack surfaces, and the difficulty of enforcing consistent protections across dynamic and resource-constrained environments. Traditional security mechanisms are not well suited to cope with these systems as they assume stable infrastructures and predictable threat models. On the other hand, Moving Target Defense (MTD) offers a proactive approach by continuously reconfiguring system parameters, reducing the attack surface and increasing uncertainty and complexity for adversaries. Nonetheless, the existing techniques for distributed environments have limited practical integration into real systems due to the complexity of systems and mechanisms. In addition, they target specific technologies or isolated mechanisms, lacking modularity and unified orchestration strategies. In this paper we present AMTD-EC, a novel adaptive MTD framework tailored to distributed infrastructures built upon container-based technologies following a threat modeling methodology. The framework leverages specific monitoring rules to enable automated detection of attacks, and includes automated selection and enforcement of adaptive response to ongoing attacks by the activation of suitable MTD mechanisms from a pool. By abstracting and orchestrating heterogeneous defense mechanisms within a unified architecture, AMTD-EC enables the scalable deployment and streamlined management of security measures across distributed Edge-Cloud environments. The experimental evaluation conducted on a Kubernetes-based testbed shows low mitigation latency and negligible system overhead when simulating realistic attack scenarios, enhancing system resiliency against dynamic threats.
Post-Quantum Cryptography (PQC) standardization efforts have led to the selection of multiple algorithms across different international organizations. While these schemes appear diverse at the algorithmic level, their security ultimately relies on a limited set of underlying hardness assumptions. In this work, we propose a graph-based representation of PQC standards that explicitly links standardized algorithms to their computational problems and captures structural relationships between these assumptions. Using network analysis techniques, we study the organization of the resulting dependency graph. Modularity analysis reveals a strongly structured ecosystem, with a densely connected algebraic core and distinct peripheral components corresponding to alternative paradigms such as hash-based and MPC-in-the-head constructions. Betweenness centrality further identifies a small number of core problems, in particular, the Shortest Vector Problem (SVP), as critical connectors within the network. Extending the analysis across standardization bodies, we observe a significant convergence toward lattice-based constructions, resulting in a concentration of dependencies on closely related hardness assumptions. This structural concentration suggests potential systemic risks, as advances in cryptanalysis could simultaneously affect multiple schemes and standards. In general, our results highlight the importance of considering structural diversity at the level of underlying assumptions in the design and evaluation of PQC ecosystems.
Neural networks based machine learning models are used in many classification problems yet privacy issues in model usage are mostly unresolved. When a model owner provides their model to other parties for inference, either the model owner has to share the parameters of the model or the model user must share their query and result with the model owner. Depending on the usage scenario, the user’s query or query result could contain sensitive information that the model provider should not see. And on the other hand, the model parameters themselves could be a valuable business secret or leak private training data. In such cases, neural network inference must be performed in a privacy preserving manner, for example, using secure multi-party computation (MPC). We propose protocols for the private inference of neural networks and evaluate them within the Sharemind MPC secure computation framework. Our protocols are based on the per-tensor quantisation scheme in PyTorch which maps floating point model parameters to 8-bit integer values. Quantisation improves the efficiency of our protocols and our protocols are interoperable with models trained using PyTorch. We extend constant round integer convolution and matrix multiplication protocols for quantised model weights, propose constant round protocols for re-quantising values and analyse the optimal secret share bit width for efficient re-quantisation. The resulting protocols are used to evaluate the VGG16 image classification network with an amortised throughput of 24.3 s per image classification in a LAN setting.
Capture-the-Flag (CTF) challenges are widely used in computer science education to support hands-on learning, problem-solving, and student engagement. However, the design of educational CTF challenges is often experience-driven and lacks systematic methodological support, particularly in higher education contexts with heterogeneous student cohorts and varying prerequisite knowledge. This paper proposes a structured method for designing CTF challenges in higher education. The approach supports the alignment of challenges with prerequisite knowledge from different computer science domains and introduces a systematic approach for calibrating and controlling challenge difficulty in heterogeneous student cohorts. The proposed method aims to facilitate curriculum-aligned and pedagogically grounded chal-lenge-based learning (CBL) experiences.
Security monitoring of edge-IoT fleets faces three structural challenges. (i) A per-node monitor is cheap but cannot see attacks that coordinate across devices. (ii) A cloud monitor sees the full fleet but pays for that view in bandwidth. (iii) Even at the cloud, a monitor built on a single RV engine can be fooled by an attacker who compromises a device, raises one malicious request, and then goes silent: once the events stop, an event-triggered monitor has nothing left to evaluate. We propose a three-layer hierarchical runtime-verification framework that addresses all three. The edge layer classifies events as they happen, the gateway layer aggregates short windows of per-device behaviour, and the cloud layer runs two complementary RV engines. MonPoly handles first-order temporal correlation over the merged alert stream: coordinated overflow (which genuinely quantifies across devices) plus per-device multi-vector APT, escalation, and persistent-campaign patterns. RTLola handles a time-triggered silent-node property that an event-triggered engine cannot detect within a bounded delay under fleet silence. We evaluate the framework on a 15-actor Docker testbed covering eight attack profiles plus a silent-bypass scenario. In the controlled labelled testbed, every device-attributable incident the framework raises names an attacker-labelled device, and the RTLola tier catches silent-bypass attempts the event-triggered tier misses. Per-event monitoring stays in the microsecond range at the edge and gateway, with low end-to-end alert-to-incident latency at the cloud.
Generative AI has rapidly reshaped products and services across multiple sectors, introducing significant risks including embedded bias, unlawful processing of personal data, and potential erosion of fundamental rights. In response, the European Union has established a complex regulatory ecosystem comprising the AIA, the DA, the DGA, and the GDPR, whose combined operationalization presents considerable challenges for stakeholders such as providers, deployers, auditors, and end-users. This paper proposes a multi-stakeholder framework that provides the structure, tooling, and secure information registration mechanisms to facilitate regulatory adherence and auditability for high-risk AI systems under the EU AI Act. The framework coordinates multiple actors through structured roles and responsibilities, automatically recording all relevant actions and decisions in a tamper-evident and verifiable manner, so that all necessary evidence is readily available in case a compliance assessment or audit is required. This is supported through smart-contract-based mechanisms for model registration, licensing, expert oversight, and decision-level traceability. Each AI prediction generates a Decision Output Object enriched with explainability metrics and optional expert validation, reinforcing transparency and human-in-the-loop supervision. Sensitive components are protected via a digital-envelope encryption scheme enabling fine-grained, role-based access control, while long-term integrity is ensured by anchoring critical artifacts on the blockchain and distributing metadata through IPFS.
Unmanned aerial vehicles are more and more often being deployed in both civilian and military contexts. One of the pressing issues of the technology is their resilience to network intrusions. This paper addresses issues related to the growing problem of cyberattacks on unmanned aerial vehicles (UAV). This paper presents an empirical evaluation of a novel benchmark dataset, UAVIDS-2025, and its suitability for training machine learning-based network intrusion detection systems. The authors used this dataset to train three different artificial intelligence models: Random Forest, XGBoost, and Deep Neural Network, trained with minimal processing and default hyperparameters. All three models achieve strong classification performance without any additional feature engineering. The findings confirm initial machine learning viability for detection of cyberattacks targeting UAV.
Complex Event Processing (CEP) enables real-time processing and analysis of large event streams across various Internet of Things (IoT) applications. Despite the deployment of CEP systems in areas that handle sensitive data, privacy protection in CEP systems has remained largely unexplored in the literature. In particular, existing approaches assume a trusted CEP system and focus mainly on pattern-level privacy, often neglecting event-level privacy. Order-revealing encryption enables comparison-based pattern detection over encrypted events without decryption, making it suitable for CEP systems, but it remains vulnerable to inference attacks from order leakage. In this paper, we propose a distributed privacy-preserving CEP framework that applies differential privacy before order-revealing encryption to reduce inference leakage. The proposed scheme achieves both event-level and pattern-level privacy against passive external adversaries and honest-but-curious internal adversaries. We formally analyze the privacy guarantees of our approach under the defined threat model and demonstrate that it achieves high detection accuracy under privacy protection and improves scalability by reducing per-node computational burden compared to a centralized CEP system.
Security operations in software-defined and virtualized networks often require coordinating heterogeneous tools for monitoring, evidence collection, and response. However, these tools expose different configuration models, APIs, and execution semantics, which complicates automation and incident response workflows. This paper presents an OpenC2-based orchestration layer that extends command-level cyber-defense coordination beyond classical enforcement actions. We introduce three actuator profiles for network-flow monitoring, file/log collection, and constrained remote command execution, and implement them using nProbe, Packetbeat, Filebeat, and Linux-based actuators. The resulting prototype provides a uniform command interface for activating monitoring functions, collecting operational evidence, and executing controlled response actions. We validate the approach through schema-conformance tests, positive and negative command cases, safety checks, and latency measurements in representative orchestration workflows.
5G Standalone (SA) networks introduce the Subscriber Concealed Identifier (SUCI) and the 5G Authentication and Key Agreement (5G-AKA) protocol to eliminate the plaintext identity exposure of earlier generations. However, privacy leakage can persist through observable protocol behaviour rather than exposed identifiers. This paper presents an experimental evaluation of two such attack classes—the SUCI-catcher and the SQN inference attack—implemented as a transparent N2-interface proxy on a fully virtualised, 3GPP Release 16-compliant testbed built on OpenAirInterface, UERANSIM, and a custom NGAP proxy. The SUCI-catcher exploits the binary outcome of the 5G-AKA challenge-response exchange as a presence oracle, successfully recording SUCIs from six subscribers and confirming target presence without recovering any keying material, weakening location privacy and unlinkability in a controlled virtualised environment. The SQN inference attack recovers subscriber sequence-number bits through controlled synchronisation failures, accurately inferring 8 bits of SQN verified against ground-truth core network state, and demonstrating that SQN changes between observations track subscriber activity, weakening undetectability in a controlled virtualised environment. Both attacks share a structural dependency on abusing core-network authentication-vector generation. Operator-level rate-limiting and anomaly detection are identified as near-term mitigations; long-term protection requires protocol-level changes to 5G-AKA.
DICOM and PACS infrastructures are essential components of modern healthcare systems but remain exposed to protocol-level security and robustness issues. Traditional fuzzing approaches are often not well suited for testing complex and stateful medical communication protocols. In this paper, we present an extension of NetworkFuzzer, a smart network protocol fuzzer, for intelligent security assessment of DICOM-based systems. Our approach combines rule-based replay, GAN-based traffic generation, LLM-assisted rule synthesis, and a reinforcement learning-guided fuzzing engine that adapts mutation strategies according to runtime feedback from the target system. We also develop an autonomous pentesting agent that integrates reconnaissance, analysis, fuzzing, and reporting capabilities to support scalable security assessments of healthcare infrastructures. Experimental evaluation on multiple real-world DICOM server implementations demonstrates the effectiveness of the proposed framework in exploring deep protocol states and identifying robustness and security issues, including hangs, crashes, and abnormal behaviors.
Weakly supervised and few-shot Large Language Models (LLMs) show great promise in handling complex text analysis without the need for extensive retraining or finetuning. In this paper, we assess practical capabilities of LLMs in disinformation analysis during the eastern border migration crisis. In particular, we evaluated and compared four specific models: Llama 3.1, Qra, Bielik, and PLLuM. Each model was tasked with labeling news articles based on credibility, disinformation types, and thematic categories. By using Cohen’s Kappa coefficient, we measured how much these models actually agree with each other. Our results show a major sensitivity gap. Particularly, while global models like Llama often view reports as neutral, the Polish-centric models flagged significantly more content as propaganda. In this paper, we investigate the origins of this phenomenon and provide details of the conducted experiments.
