Log messages can play a significant role in different activities, including debugging, monitoring, and security analysis. Yet, they are often placed only for the convenience of development rather than for security relevance. For reliable logging to support security and forensic investigations, logs should be positioned along all execution paths that lead to critical operations that process untrusted user input. This placement ensures that any such operation leaves a reliable and observable trace for forensic trail. In this article, we present BlindSpot, an automated static analysis tool to extract inter-procedural paths where the parameters of critical calls appear within a log statement. Our approach incorporates dataflow and control-flow analysis to precisely capture the relationships between program nodes and model those log elements that are consistently encountered from input to untrusted critical calls (pre-execution), as well as those that always follow afterward (post-execution). We tested our approach on 10 desktop applications, containing 124 sensitive operations that process untrusted input. Our tool revealed significant inconsistencies in logging practices: only 37 of these actions (29.8%) have any associated log statements, and just 8 are logged prior to execution. This reliance on post-execution logging presents a risk because if a critical operation is exploited or causes a crash or failure, the log statements that follow may never execute, leaving no trace for forensic analysis. We also find that log statements are very often conditional and executed only along specific paths. In our analysis, we found that only 5 out of 124 critical operations (4%) are associated with an unconditional log. Through BlindSpot, we demonstrate how the inconsistency, conditionality, and absence of logs around critical operations raise concerns about the reliability of logs as a source of evidence.
Operational technology (OT) is foundational to critical sectors such as water, energy and manufacturing, yet securing it presents unique challenges. Conventional IT penetration testing methods cannot be directly applied to OT because of the risks of disrupting live production environments. As a result, there is little consensus on how penetration testing should be conducted in OT environments, and practice remains inconsistent and poorly understood. This article explores the state of OT penetration testing through a review of existing literature and qualitative interviews with both practitioners and procurers. The findings highlight three central challenges: the absence of standardised methodologies, the limited relevance of IT-style vulnerability reporting, and cultural tensions between cyber security specialists and operational engineers. From these insights, the article proposes a set of short-, medium- and long-term recommendations to improve practice and points to key areas for future research.
The Internet of Medical Things (IoMT) has transformed healthcare delivery through medical devices, remote patient monitoring, and real-time clinical decision support. However, the proliferation of IoMT devices introduces security vulnerabilities that put patient safety and data privacy at risk. Intrusion Detection Systems (IDS) have emerged as essential components for protecting IoMT networks from cyberattacks. This article presents a systematic review of IoMT-IDS research, analyzing 53 high-quality papers published between 2020 and 2025, identified through database searches spanning 2016–2025 across IEEE Xplore, Springer, ScienceDirect, and ACM Digital Library. We organize the literature through a comprehensive taxonomy spanning classical machine learning approaches, deep learning architectures, transformer-based models, federated learning frameworks, and anomaly detection techniques. Our analysis reveals significant diversity across detection approaches and impressive within-dataset performance achievements, yet identifies persistent gaps in evaluation practices. While detection performance is consistently reported across studies, four other evaluation dimensions remain unaddressed: computational efficiency profiling, explainability integration, standardized benchmarking practices, and cross-dataset generalization assessment. We synthesize the literature into these five evaluation dimensions as a foundation for comprehensive IoMT-IDS assessment. We then analyze state-of-the-art approaches, highlight tradeoffs between competing requirements, and chart future research directions toward unified evaluation methodologies that support deployment in safety-critical healthcare environments.
Understanding cyber attacker behavior is a fundamental requirement for advancing effective threat detection and response mechanisms. Nevertheless, existing cyber attacker profiling techniques often lack sufficient behavioral granularity to reliably distinguish adversaries, particularly when analyzing highly variable, noisy, and chained command sequences. To capture true operational semantics, this paper presents a novel behavior-driven profiling framework based on empirical data collected from SSH honeypot deployments. The proposed methodology systematically parses these complex command chains to identify distinct behavioral patterns and categorize attackers accordingly. A total of 176 unique attack patterns were extracted and organized into 18 behavioral clusters, which were subsequently mapped to the MITRE ATT&CK framework for structured interpretation. Cluster validity and reliability were evaluated using Indicators of Compromise (IoCs) and statistical validation methods. The evaluation confirmed seven robust behavioral clusters, while five additional clusters exhibited strong alignment with known IoCs, demonstrating the effectiveness, robustness, and practical applicability of the proposed profiling approach for real-world cyber threat analysis.
Social media platforms have become valuable sources for detecting phishing due to their widespread use and rapid content sharing. Among them, Telegram stands out as an open, data-rich environment, offering access to large public groups where phishing links frequently appear. This study investigates whether Large Language Models (LLMs) can serve as effective tools for phishing detection on Telegram. We systematically evaluate three representative LLMs—GPT-4o, LLaMA-3.1, and DeepSeek V3—across multiple prompting strategies and input configurations. For comparison, we employ an improved version of traditional feature-based machine learning models from the literature, enhanced with resampling techniques to address the strong class imbalance in our data. Experimental results show that while LLMs demonstrate some potential, achieving a Recall for the phishing class of up to 0.554 with DeepSeek—representing the model’s ability to correctly identify 55.4% of all actual phishing messages—they suffer from extremely low precision (0.065 with DeepSeek), resulting in a low F1-score of 0.114. In contrast, our feature-based models achieve a significantly more reliable precision–recall balance, with an overall phishing F1-score of 0.749 and a precision of 0.738. The findings in this study highlight the potential and current limitations of LLMs for phishing detection in dynamic instant messaging environments and emphasize the superior performance of platform-tailored models.
Federated learning based solutions enable multiple organizations to build a machine learning model using the collective knowledge of the organizations without exchanging the sensitive raw data of any participant. The key roadblock for widespread adoption is the possibility of training data leakage in federated learning models, which have been reported in federated learning models for image processing, natural language processing, and tabular data. However, the impact and feasibility of such privacy leakage attacks on federated learning-based intrusion detection systems (IDS) remains largely unexplored. In this work, for the first time, we focus on the problem of training data privacy in federated learning based network intrusion detection systems by exploring the implementations, techniques and limitations of existing federated IDS solutions. First, we provide a systematic study of the landscape of machine learning based network intrusion detection systems and federated learning based network intrusion detection systems. Second, we demonstrate that data leakage attacks on federated learning based intrusion detection systems are yet to be explored to the full extent by the research community. Finally, we describe the existing defenses against such attacks and suggest the possibility of using these approaches for federated network intrusion detection systems in the future.
High-privilege insiders, such as IT administrators, pose a significant threat to data integrity, whether through accidental errors or malicious intent. Unauthorized modifications or deletions of critical data—such as system logs and access records—can cause severe operational and security disruptions. Traditionally, these threats are mitigated using hardware-based solutions such as Write-Once Read-Many (WORM) storage. While effective, these approaches have notable drawbacks, including high deployment costs and irreversible consumption of storage blocks, which cannot be reused once written. The goal of this article is to explore purely software-based solutions to these challenges. To this end, we design VaultFS, a software-only file system for Linux environments tailored for the protection of cold data—data that must remain accessible but unmodifiable. VaultFS enables writing through a standard file system interface while ensuring strict immutability and undeletability for a pre-defined time window, potentially infinite. Even threads operating with (effective)root-id permissions cannot alter or remove stored data, preserving integrity against privilege escalation threats. Achieving this requires addressing a number of key technical challenges, such as ensuring a trusted time reference to enforce temporal protection guarantees, and forcing a suited admission control policy to avoid any interaction with the file system that could be potentially dangerous, including access to the file system block device via common services offered by the Linux virtual file system. Furthermore, VaulFS mitigates storage-exhaustion Denial-of-Service (DoS) attacks, where untrusted applications flood the file system with non-removable content. We evaluate the integration of VaultFS with practical applications, demonstrating full compatibility with all seven backup tools assessed—including mysqldump and rsynch —and seamless integration with most of the video surveillance applications tested—including ivideon and ZoneMinder . Furthermore, our experimental evaluation shows that VaultFS incurs only a 7–12% performance overhead compared to the common Linux Ext4 file system under the usage of read/write intensive applications performing file-copy operations.
Cybersecurity is increasingly in the crosshairs of policy makers, as evidenced by the introduction of far-reaching legal frameworks around the globe. One concrete example of cybersecurity policy is how to deal with vulnerability disclosures. Organisations increasingly introduce vulnerability disclosure policies and in some cases public sector bodies are even required by law to have such policies. In this work we study the effects of these policies in practice. Using the process specified on an organisation’s website, or absent such a process following community best practices, we disclose an email vulnerability affecting a large number of organisations. This vulnerability allows arbitrary actors to send mail on behalf of affected organisations by abusing shared infrastructure. Our disclosure campaign focuses specifically on public and critical infrastructure organisations that are required by law to handle such disclosures. We find that having a policy makes it easier to contact organisations regarding security vulnerabilities. Nevertheless, even with a policy in place, over half of our reports remain unanswered and unresolved after 90 days. Based on our findings, we provide recommendations to policy makers and organisations how to better shape their vulnerability disclosure processes.
Human-Operated Ransomware (HoR) is one of the most persistent and evolving threats in cybersecurity, as attackers use changing Tactics, Techniques, and Procedures (TTPs) to evade traditional detection. The lack of structured and publicly available TTP-level datasets has limited the development of models capable of identifying HoR behavior early. In this study, we construct a dataset of TTP sequences from 15 prominent ransomware families observed in 2023 and 2024, structured according to the MITRE ATT&CK framework. We evaluate a range of sequence modeling approaches, including Markov chains, n-gram analysis, LSTM, GRU, and RNN, to classify ransomware behavior based on the progression of observed TTPs. The RNN model achieved the highest accuracy of 82% and an AUC of 0.9694(95% CI: 0.0087-0.0168), with an average false positive rate between 0.0087 and 0.0168 using 10-fold cross-validation. SMOTE was used to address class imbalance, improving model accuracy by 6% (from 76% to 82%). These results show that learning from ordered TTP patterns can support timely detection of ransomware activity, enabling earlier intervention in threat response workflows.
Financial crime is increasingly facilitated by technology and globalization, demanding advanced IT tools for detection. The Fraud Detection System proposed in this article is devised to operate under pragmatic operational constraints inherent to financial institutions, such as extreme class imbalance, due to the rarity of fraudulent events, or the continuously evolving fraud patterns (concept drift) driven by adversarial adaptation, and, most significant, the stochastic delays in obtaining verified feedback, crucial for model supervision. This research develops financial transaction monitoring in a data stream context; it details a developed streaming Machine Learning (ML) pipeline, designed with a lightweight yet powerful Data Stream Management System (DSMS) for real-time feature engineering, and a multi-stage analytical engine that orchestrates diverse detection logic. This engine integrates deterministic rules with adaptive ML models and is coupled with a dynamic decision threshold management system to optimize the precision-recall tradeoff under operational pressures. A core contribution is the systematic, empirical comparison of diverse adaptive learning strategies, ranging from instance-incremental to various batch-incremental methods, to assess their adaptability and effectiveness under these operational conditions. Furthermore, the article describes an actionable interpretability framework designed to synthesize low-level feature attributions into user-centric concept importances, enhancing the utilities at the disposal of investigators.
The rapid rise of deepfake technology continues to challenge digital security, trust, and misinformation control particularly for celebrities and public figures, whose identities are frequently exploited. This article introduces a novel dual paradigm deepfake detection framework that integrates a classical attention-enhanced EfficientNetB4 model with a Quantum Trained Convolutional Neural Network (QT-CNN). The classical stage leverages spatial attention and siamese feature alignment to highlight manipulation sensitive facial regions and improve cross-dataset generalization. Building on this, the QT-CNN employs parameterized quantum circuits and quantum-to-classical parameter mapping to reduce model complexity while preserving detection accuracy. Comprehensive experiments on a large-scale South Asian celebrity dataset, an underrepresented demographic in existing benchmarks alongside FF++ and DFDC, demonstrate that the hybrid approach achieves robust performance, including 94.5% accuracy on in-distribution data and strong generalization under demographic, corruption, and compression shifts. The QT-CNN further reduces trainable parameters by nearly 70%, suggesting a promising pathway for efficient deployment in resource-constrained, high-volume environments such as social media moderation pipelines. This work contributes a scalable, demographically inclusive, and quantum informed methodology toward securing digital ecosystems in both current and emerging post quantum environments.
Security conferences are important venues for information sharing, where academics and practitioners share knowledge about new attacks and state-of-the-art defenses. Despite their importance, researchers have not systematically examined who shares information and which security topics are discussed. To address this gap, our article characterizes the speakers, sponsors, and topics presented at prestigious academic and industry security conferences. We compile a longitudinal dataset containing 9,728 abstracts and 1,686 sponsors across 4 academic and 6 industry conferences. Our findings show limited information sharing between industry and academia. Conferences vary significantly in how equitably talks and authorship are distributed across individuals. The topics of academic and industry abstracts display consistent coverage of techniques within the MITRE ATT&CK framework. Top-tier academic conferences, as well as DEFCON and Black Hat, address the governance, response, and recovery functions of the NIST Cybersecurity Framework inconsistently. Commercial information security and insurance conferences (RSA, Gartner, Advisen, and NetDiligence) more consistently cover the framework. Prevention and detection were the most common topics in the sample period, with no clear temporal trends.
Computing side-channel research explores the manner in which physical emanations from systems can be used to reconstruct data. Acoustic side-channels are those physical emanations that produce a sonic frequency that is subsonic, supersonic, or considered in the range of human hearing [12]. Acoustic Side-Channel Attacks (SCAs) are typically performed passively: a listening device captures aural frequencies from a machine via a microphone that are transmitted to the attacker for analysis [8, 12, 14]. Machine learning models have been presented to classify individual keystrokes according to variations in acoustic frequency [2]. Furthermore, the SonarSnoop framework presents a novel active approach that involves both generating and recording aural frequencies acting as a type of sonar system to record physical motion [7]. This research attempts to develop a supervised machine learning model to classify finger motion to collect login credentials typed on a laptop keyboard. The active acoustic side-channel has been used to track two-dimensional finger motion, but three-dimensional finger tracking using active acoustics is novel. The model as trained in this study incorrectly inferred labels on unseen data; however, we found and demonstrated that training with more samples per label may result in greater success during inference.
The financial impact of fraudulent e-commerce schemes has been increasing steadily. Several research reports demonstrate that some threat actors compromise legitimate web sites and deploy malware for black-hat Search Engine Optimization (SEO). This malware facilitates SEO poisoning, causing search engines to display deceptive lure pages as if hosted on the compromised sites, effectively redirecting users to fraudulent e-commerce platforms and increasing the risk of victimization. This study focuses on these threat actors and their tactics. To investigate relationships between malware families employed by these groups, we collected data on 2,852 command and control (C2) servers associated with 10 distinct malware families, alongside 697,816 fake e-commerce sites identified through these servers. We subsequently analyzed these data using Maltego, a widely recognized link analysis tool. Our results suggest the presence of four distinct groups each utilizing a single, unique malware family, and two groups operating multiple families. This analysis also provides valuable insights into the characteristics of these malware families.
Software supply-chain security requires provenance mechanisms that support reproducibility and vulnerability assessment under dynamic execution conditions. Conventional Software Bills of Materials (SBOMs) provide static dependency inventories but cannot capture runtime behaviour, environment drift, or exploitability context. This paper introduces agentic Artificial Intelligence Bills of Materials (AIBOMs), extending SBOMs into active provenance artefacts through autonomous, policy-constrained reasoning. We present an agentic AIBOM framework based on a multi-agent architecture comprising (i) a baseline environment reconstruction agent (MCP), (ii) a runtime dependency and drift-monitoring agent (A2A), and (iii) a policy-aware vulnerability and VEX reasoning agent (AGNTCY). These agents generate contextual exploitability assertions by combining runtime execution evidence, dependency usage, and environmental mitigations with ISO/IEC 20153:2025 Common Security Advisory Framework (CSAF) v2.0 semantics. Exploitability is expressed via structured VEX assertions rather than enforcement actions. The framework introduces minimal, standards-aligned schema extensions to CycloneDX and SPDX, capturing execution context, dependency evolution, and agent decision provenance while preserving interoperability. Evaluation across heterogeneous analytical workloads demonstrates improved runtime dependency capture, reproducibility fidelity, and stability of vulnerability interpretation compared with established provenance systems, with low computational overhead. Ablation studies confirm that each agent contributes distinct capabilities unavailable through deterministic automation.
The Internet of Things (IoT) has become a cornerstone in modern automation and data exchange, with IoT devices increasingly embedded into daily life. This development coincides with a rapid growth in mobile device usage and the proliferation of interactive mobile applications, catalyzing the evolution of mobile crowdsourcing. Such applications offer a range of functionalities, from automated data collection through sensor driven and location aware services to manual input via user surveys and feedback. For mobile crowdsourcing, particularly in anonymous and ever changing environments, modern trust management systems rarely deal with the problem of participants credibility and reliability. This paper presents a hybrid approach that integrates sophisticated trust management techniques with Support Vector Machine (SVM) to improve the security of mobile crowdsourcing platforms. The system considerably enhances the reliability of trust evaluations, successfully protecting against malicious actors. It provides a thorough trustworthiness score to contributors by utilizing a variety of variables, including social networking site data, reputation measures, and user behavior patterns. The high efficacy of our model is demonstrated by an exploratory evaluation of a real mobile crowdsourcing platform, which achieved an accuracy rate of approximately 99.85%.
Vulnerability disclosure is the practice of a finder disclosing a newly found vulnerability to vendors. It has received best practices to ensure communication between stakeholders. However, the practice of a finder or vendor notifying end-users about vulnerable systems and mitigation plans has not received the same attention and guidelines for performing it at scale. We identify the practice as vulnerability notification, which shares similarities with disclosure but presents other challenges and requires different approaches. In vulnerability notification, a finder targets known vulnerabilities or misconfigurations using active scans or datasets to determine how many systems or services remain vulnerable. The scale and complexity of vulnerability notification to end-users are often significantly greater than those of multi-party disclosure to vendors. These place an increasing burden on finders to inform stakeholders on time, especially for academic security researchers, ethical hackers, and practitioners. Based on our experience with notifications and academic publications documenting disclosure and notification operations, we conduct a meta-review of how researchers have adopted best practices, pursued different strategies, and reflected on their operations over the years. Drawing on the meta-review and suggestions from security communities, we propose new best practices for finders to perform vulnerability disclosure, particularly vulnerability notification at scale.
Adversary emulation tools facilitate scripting and automated execution of cyber attack chains, thereby reducing costs and manual expert effort required for security testing, cyber exercises, and intrusion detection research. However, due to the fact that existing tools typically rely on agents installed on target systems, they leave suspicious traces that make it easy to distinguish their activities from those of real human attackers. Moreover, these tools often lack relevant capabilities, such as handling of interactive prompts, and are unsuitable for emulating specific stages of the kill chain, such as initial access. This paper thus introduces AttackMate, an open-source attack scripting language and execution engine designed to mimic behavior patterns of actual attackers. We validate the tool in a case study covering common attack steps including privilege escalation, information gathering, and lateral movement. Our results indicate that log artifacts resulting from AttackMate's activities resemble those produced by human attackers more closely than those generated by standard adversary emulation tools.
The emulation of multi-step attacks attributed to advanced persistent threats is valuable for training defenders and evaluating defense tools. In this paper, we discuss the numerous challenges and desired attributes associated with such automation. Additionally, we introduce the use of Effects Language (EL), a visual programming language with graph-based operational semantics, as a solution to address many of these challenges and requirements. We formally define the execution semantics of EL, and prove important execution properties. Furthermore, we showcase the application of EL to codify attacks using an example from one of the publicly available attack scenarios. We also demonstrate how EL can be utilized to provide proof-of-attack of complex multi-step attacks. Our results highlight the improvements in time and resource efficiency achieved through the use of EL for repeatable automation.
The use of multi-threading and file prioritization methods has accelerated the speed at which ransomware encrypts files. To minimize file loss during the ransomware attack, detecting file modifications at the earliest execution stage is considered very important. To achieve this, selecting files as traps and monitoring changes to them is a practical way to deal with modern ransomware variants. This approach minimizes overhead on the endpoint, facilitating early identification of ransomware. This paper evaluates various machine learning-based trap selection methods for reducing file loss, detection delay, and endpoint overhead. We specifically examine non-parametric clustering methods such as Affinity Propagation, Gaussian Mixture Models, Mean Shift, and Optics to assess their effectiveness in trap selection for ransomware detection. These methods select M files from a directory with N files (M<N) and use them as traps. In order to address the shortcomings of existing machine learning-based trap selection methods, we propose APFO (Affinity Propagation with File Order). This method is an improvement upon existing non-parametric clustering-based trap selection methods, and it helps to reduce the amount of file loss and detection delay encountered. APFO demonstrates a minimal file loss percentage of 0.32 contemporary ransomware variants, including rapid encryption variants of lock-bit, AvosLocker, and Babuk.
