Digital fatigue is increasing due to ongoing digitization, yet understanding digital processes remains crucial. Cybercriminals exploit the lack of knowledge and awareness among users. To address this, we developed “Hook & Match”, a Pair Matching Phishing Card game, which combines the familiar non-digital principle of pair matching with phishing education to avoid contributing to digital fatigue. In “Hook & Match”, players match cards to learn how to distinguish different phishing attempts from legitimate examples. Our evaluation shows that playing “Hook & Match” significantly improves users’ ability to identify phishing compared with legitimate cases. We also investigated whether prior knowledge had an impact on the learning outcomes and found that, regardless of their initial level of expertise, both knowledgeable users and novices benefited from playing the game. However, participants with the lowest initial level benefited the most. This demonstrates that “Hook & Match” is an effective tool for phishing education without exacerbating digital fatigue.
With the prosperity of the Internet of Medical Things (IoMT), ensuring privacy regulation-compliant, patient-trusted data management without reliance on IoMT providers is critical. While legislation mandates providers to be accountable, the trustworthiness of IoMT providers is undermined by their conflicting roles of data processor and data processing activity logger. A curious or malicious provider can easily copy and abuse data undetected, with both roles. Moreover, the IoMT data processing code is a proprietary asset of the provider, making public verification infeasible. This necessitates a trusted data processing paradigm that operates securely even when both the IoMT provider and their code are untrusted. By leveraging trusted execution environments (TEE), we propose a time-series data management system (DMS) with trust shifted from IoMT providers to hardware vendors, enabling low-overhead trusted data processing with a small trusted computing base (TCB). Benchmarks are performed over a minimal prototype, and the results show low latency, low memory footprint, and a small TCB for custom code (excluding dependencies). Security analysis of the architecture confirms end-to-end confidentiality, integrity, and rollback resistance under a strong threat model with untrusted providers. Our approach complies with privacy principles, is scalable via patient-wise microservice partitioning for edge-cloud deployments, and can integrate cross-platform TEE solutions for portability.
With the rapid advancement of artificial intelligence (AI), particularly the widespread adoption of large language models (LLMs) in code comprehension and analysis, their strong semantic parsing capabilities have introduced new threats to software security. Attackers can exploit LLMs to reverse-engineer the deeper semantic logic of code, steal core algorithms, or uncover vulnerabilities, thereby endangering software intellectual property and system security. This work introduces SemAder , a structure–semantics joint induction framework that generates adversarial yet function-preserving binaries to mislead LLMs’ functional judgments in binary analysis, thereby reducing the reliability of LLM-assisted semantic analysis during reverse engineering. SemAder comprises three core components: (1) a control-flow-labeled induced corpus annotated with structural tags and code semantics; (2) a hybrid similarity-driven corpus selection mechanism that favors structural proximity with semantic divergence; and (3) a reinforcement-learning-driven semantic fusion pipeline that incorporates constant externalization and context-aware semantic enhancement to strengthen induction against high-capability LLMs. Experimental results across eight LLM evaluators demonstrate that SemAder consistently shifts model predictions toward the induced target category, achieving an average induction gap of 0.77 and maintaining effectiveness under adversarial prompt variants and multi-agent post-processing workflows. SemAder also misleads the CLAP code classification model (-63.5% original-class confidence) and reduces similarity scores across four binary similarity detectors (Asm2Vec, BinDiff, SAFE, Gemini) to an average of 0.51, with only 12.7% average binary size increase and 8.8% average runtime overhead.
The rapid growth of Decentralized Finance (DeFi) has been accompanied by increasingly sophisticated security threats. Price Oracle Manipulation Attacks (POMA), a critical vulnerability, have evolved beyond simple economic exploits to include complex, multi-transaction attacks that exploit smart contract logic, causing hundreds of millions in losses. State-of-the-art detection methods, however, often focus on single-transaction, economic manipulations and typically fail to identify these emerging attack vectors, particularly when smart contract source code is unavailable. This article introduces a novel, EVM-compatible detection pipeline that addresses this gap. By combining transaction event logs and execution traces, we engineer a rich set of semantic and structural features that capture the underlying behavior of on-chain operations. We train a regularized autoencoder exclusively on the features of benign transactions to learn a deep representation of normal activity, flagging significant deviations as malicious. Our evaluation demonstrates the effectiveness of this approach, achieving 100% recall on a comprehensive dataset of single-transaction attacks and 98.25% event-level recall on a new, manually collected dataset of real-world multi-transaction exploits, with an overall precision of 97.15%. We present a robust, learning-based model capable of identifying both known and unseen POMA variants without relying on source code. Furthermore, we contribute a new dataset of multi-transaction attacks to foster further research, providing a more generalizable and resilient approach to securing the DeFi ecosystem.
Deep neural networks for image classification require protection against unauthorized use and redistribution. Existing watermarking methods suffer from a critical vulnerability: watermarks are always active and detectable, allowing adversaries to identify and remove them before deployment. We propose DormMark, a novel framework for image classification models that introduces delayed-activation watermarks which remain dormant and hidden under deployment-time black-box query auditing during initial deployment, but automatically activate upon fine-tuning. Our approach employs a three-stage training paradigm: (1) embedding watermarks using triggered samples, (2) masking to suppress watermark functionality while preserving its latent presence, and (3) activation through standard fine-tuning without owner intervention. This mechanism exploits neural networks’ forgetting-remembering behaviors during continued training, creating a fragile equilibrium that behaves similarly to a clean model under deployment-time black-box auditing but reliably manifests ownership indicators after modification. We consider a private-key black-box verification setting in which the owner keeps the concrete trigger instances secret. Experiments across multiple architectures (VGG19, ResNet-18/56, DenseNet-121, WideResNet-34) and datasets (CIFAR-10, CIFAR-100, GTSRB) demonstrate 100% watermark success rates, high imperceptibility (PSNR > 38 dB, SSIM = 0.99), negligible accuracy loss (< 0.04%), and robustness against 80% parameter pruning. DormMark represents a paradigm shift from static to conditionally-activated ownership verification, providing a more robust framework for intellectual property protection.
Cloud computing supports the Internet of Things (IoT) in handling diverse and large-scale data. However, outsourcing data control to the cloud raises security concerns, particularly in key management. Although Ciphertext-Policy Attribute-Based Encryption (CP-ABE) preserves data confidentiality, it entrusts key management to a centralized attribute authority, resulting in the key escrow problem. Furthermore, existing CP-ABE schemes lack mechanisms for key verification and identity authentication, leaving IoT systems susceptible to key errors and impersonation attacks. To overcome these limitations, we propose Secure and User-autonomous Attribute-based Data Sharing (SUAD) for cloud-assisted IoT. The SUAD scheme transfers key management from the authority to data users themselves, thereby eliminating key escrow. Built on a data user-centric architecture, the SUAD scheme removes the decryption privilege of the attribute authority. To prevent key forgeries and operational errors, we design a correctness verification mechanism covering five critical keys and the decryption result, along with a two-way interactive authentication protocol based on the Schnorr scheme for reliable identity verification. The SUAD scheme further supports dynamic user management, enabling user logout, replacement, and joining while optimizing maintenance overhead through periodic updates. We formally prove that SUAD achieves selective IND-CCA security in the random oracle model. Both theoretical analysis and experimental evaluations demonstrate that SUAD enhances user autonomy and strengthens security without incurring additional encryption or decryption costs, confirming its practicality for IoT deployments.
Living-off-the-land (LOTL) techniques, which exploit legitimate system utilities to execute malicious commands, pose significant challenges to cyber-threat detection by blending with benign behavior. Current state-of-the-art machine learning (ML) detection methods suffer from two critical limitations: (1) a need for large-scale datasets that capture LOTL behaviors, essential for detection at low false-positive rates (FPR) and high true-positive rates (TPR), and (2) a lack of adversarial manipulation evaluations, despite the inherent presence of adaptive attackers in cybersecurity contexts. To address these challenges, we introduce a novel, cyber-security focused data synthesis (DS) framework that augments malicious LOTL samples by combining threat intelligence with legitimate baselines from enterprise networks. We evaluate our framework in a large-scale production environment, focusing on the detection of Linux LOTL reverse shells. The resulting dataset and models—collectively referred to as QuasarNix —enable ML detectors that detect roughly 60% of malicious reverse shells at an industry-grade FPR = 10 -6 , whereas non-augmented baselines remain effectively blind at this operating point. We demonstrate that unprotected ML models remain vulnerable to black-box evasion attacks. To counteract these risks, we incorporate adversarial training into our DS framework, enhancing the robustness of our LOTL detection models. Through an explainability analysis, we confirm that QuasarNix provide detection engineers with evidence-based attribution, aligning with cybersecurity domain expertise. To foster reproducibility, we publicly release our framework implementation, 1 synthesized dataset, 2 and pre-trained models. 3
The random number generation capabilities of the GNU/Linux operating system are subject to certain limitations. As of Linux version 5.6, /dev/random operates in a non-blocking manner and, as such, no longer satisfies the criteria for a True Random Number Generator (TRNG). While dedicated quantum random number generator (QRNG) hardware is the preferred source of unpredictable entropy, it is often expensive and difficult to deploy in virtualized/cloud environments and Internet of Things (IoT) devices. Furthermore, hardware RNG integration typically requires cryptographic applications to adhere to vendor-specific APIs. This article proposes a user-space integration approach for a shared , potentially remote QRNG device. We develop a QRNG service on top of D-Bus, a ubiquitous inter-process communication framework. It serves as an interface for applications to retrieve true random numbers. Communication with the remote QRNG device occurs over mutually authenticated TLS 1.3 channels, protected by post-quantum cryptography (PQC) algorithms. We show, as a proof-of-concept, how the proposed D-Bus service can be integrated into the OpenSSL 3 cryptographic library, demonstrating the use of TRNG in a wide range of Linux applications. Our approach is resistant to entropy starvation attacks, supports sharing a QRNG across host and virtualized environments, requires no kernel-level or system-wide modifications, supports mixing multiple sources of entropy, and configuration of post-processing. It provides applications with a TRNG interface suitable for information-theoretically secure (ITS) use cases.
Vehicular misbehavior detection faces multiple technical challenges, including machine learning-adaptive attacks and trust management issues. A critical concern is the problem of malicious vehicle with high-reputation, in which malicious vehicles exploit trust-based security by maintaining legitimate behavioral profiles while strategically injecting malicious content. This behavior creates systemic vulnerabilities that compromise network trust infrastructures. Attacks affecting high-reputation malicious behavior detection include both external and internal types, requiring holistic defense mechanisms. However, current vehicular security research lacks unified defense against both internal and external attacks. Typically, studies that effectively resist internal attackers demonstrate the weaker defense against external attacks, and vice versa. To mitigate these concerns, we have designed a Training Verdict Autonomous Vehicle Networks architecture (TV-AVN) that develops a novel Verdict Misbehavior Detection System (V-MDS) by combining machine learning with reputation mechanism. The proposed scheme incorporates a public key cryptosystem to enhance security during basic safety message transmission. A local authority regularly consolidates detection outcomes to update vehicle reputation scores. In comprehensive experimental comparisons, our approach demonstrates robust-level security performance, with formal verification tools validating the security robustness of our proposed mechanism. For position falsification attacks, our method achieves average detection performance of 0.99 Precision , 0.98 Recall , and 0.98 F 1- score . Moreover, the proposed approach demonstrates superior resilience against intelligent attacks involving high-reputation attackers. Although the detection performance experiences degradation, our method remains more stable than existing approaches, which suffer rapid deterioration. In summary, TV-AVN establishes reliable communication for vehicle users, maintaining long-term network quality and preserving user confidence in the system.
The burgeoning prevalence of Ethereum phishing behavior has iCSUR-2025-0155mposed substantial constraints on the advancement of blockchain finance, resulting in losses of more than $7.7 billion to date, so it is urgent to detect it in time. Currently, available detection methods usually focus on the spatial features within transaction graphs. These methods often employ shallow mining techniques on small samples. As a result, they may overlook certain aspects of interaction patterns, such as temporal behavior. Additionally, their data mining capability is limited due to the small sample sizes. In this study, we propose a graph contrastive learning framework to enrich features of accounts behavior patterns with restricted samples to overcome these limitations. Firstly, we construct an Ethereum interaction graph with the multi-graph involving more temporal information centered with labeled nodes and lighten it with our strategy. Secondly, to comprehensively characterize the accounts pattern, we design the encoder part with the GAT-LSTM model based on attention mechanism fusing statistical features , fine-grained temporal behavioral features and graph structural semantic features . Thirdly, to moderate the sparsity of phishing nodes, we employ data augmentation and contrastive learning to fully mine sparse node information. Moreover, we carried out an in-depth experimental evaluation. The CMD-EPD approach, boasting an F 1 -score of 0.87, outperformed all comparison methods. We also executed a thorough case study to analyze phishing accounts phenomenological indicators which back up the superiority of our framework.
Large Language Models (LLM) such as ChatGPT, Meta AI, and Google Gemini have become highly accessible and ubiquitous across a wide range of applications, including speech synthesis, code generation, and media content creation. Recent research indicates that an alternative motivation for such tools is to rapidly develop malware to conduct cyber attacks. In this article, we investigate how generative LLM tools can be used to synthesise cyber-attacks targeting Industrial Control Systems (ICS). We introduce a methodology that uses LLMs to generate attack techniques based on the MITRE attack framework to target a variety of Programmable Logic Controllers (PLC) models from by different industrial vendors. We investigate the capability of five leading off-the-shelf LLMs by providing different levels of attacker context to enhance the generation. Through a comprehensive evaluation of the generated code and the resulting LLM outputs, we demonstrate that current general-purpose LLMs are capable of identifying the necessary steps required to synthesise attacks that can manipulate the operations of real PLCs. We highlight that the success of LLM-generated PLC cyberattacks depends on the level of target context initially provided, emphasising the importance of mitigating early-stage reconnaissance attacks in OT environments.
In today’s digital landscape, our interactions, from professional collaborations to personal data sharing involving photos, movies, and documents, have largely moved online. While transitioning these activities to digital platforms provides considerable convenience, it poses significant challenges in efficiently managing and securely erasing shared data in compliance with privacy regulations. Digital forgetting, particularly in co-owned data, transcends being merely desirable and becomes a mandate. Conventional data management paradigms, including cryptographic erasure techniques, typically apply uniform deletion across all stakeholders, neglecting audience-specific expiration and co-owner participation in deletion, which limits their applicability in contemporary cloud storage ecosystems. This article introduces a Policy-Based Conjunctive Scheme (PBCS) that enables conjunctive decision-making for data access and collaborative data forgetting, aligning with the General Data Protection Regulation (GDPR)’s Right to be Forgotten (RTBF). PBCS allows owners to upload their data to the cloud securely and offers policy-based access control to co-owners, granting them the ability to influence decisions about data deletion via democratic voting mechanisms significantly. The scheme leverages conjunctive access thresholds and mechanisms that gradually make data irretrievable. By integrating cryptographic primitives and Lagrange interpolation-based decay, PBCS supports a flexible, conjunctive governance model that upholds privacy and enhances the data lifecycle. We provide a formal analysis and an experimental evaluation of our scheme.
Recent attacks on federated learning (FL) can introduce malicious model updates that can circumvent widely adopted Euclidean distance-based detection methods. This article proposes a novel defense strategy, referred to as LayerCAM-AE, designed to counteract model poisoning in FL. The LayerCAM-AE puts forth a new Layer Class Activation Mapping (LayerCAM) integrated with an autoencoder (AE), significantly enhancing detection capabilities. Specifically, LayerCAM-AE generates a heat map for each local model update, which is then transformed into a more compact visual explanation. The autoencoder processes the LayerCAM heat maps from the local model updates, improving their distinctiveness and increasing the accuracy in spotting anomalous maps and malicious local models. To mitigate the risk of misclassifications in LayerCAM-AE, a voting algorithm is developed, where a local model update is flagged as malicious if its heat maps are consistently suspicious over several communication rounds. Extensive tests on the SVHN and CIFAR-100 datasets are performed under both Independent and Identically Distributed (IID) and non-IID settings in comparison with the state-of-the-art ResNet-50 and REGNETY-800MF defense models. The experimental results show that LayerCAM-AE increases detection rates (Recall: 1.0, Precision: 1.0, FPR: 0.0, Accuracy: 1.0, F1 score: 1.0, AUC: 1.0) and the test accuracy of FL, surpassing both the ResNet-50 and REGNETY-800MF. Our code is available at: https://github.com/jjzgeeks/LayerCAM-AE .
Deep Learning (DL) has significantly advanced artificial intelligence (AI) across numerous applications. However, its foundational reliance on centralized data collection introduces critical limitations concerning privacy, security, and scalability. With the continued proliferation of the Internet of Things (IoT), massive volumes of sensitive data are generated at the network edge, necessitating collaborative learning systems capable of securely sharing information without compromising confidentiality. Federated Learning (FL) partially addresses these challenges by enabling on-device model training, but its dependence on a central coordinator remains a core vulnerability, creating communication bottlenecks, fairness constraints, and susceptibility to single-point-of-failure (SPOF) risks. Swarm Learning (SL), developed in collaboration with Hewlett Packard Enterprise (HPE), represents a decentralized advancement that mitigates these limitations by eliminating central orchestration. Through the integration of blockchain technology with distributed machine learning (ML), SL establishes a secure, transparent, and fault-tolerant paradigm for peer-to-peer model exchange and aggregation. This survey provides a comprehensive overview of the architectural foundations, enabling technologies, and applications of SL in sensitive domains, including healthcare, the Internet of Vehicles (IoV), and Industrial IoT. In addition, it introduces a comparative taxonomy that systematically categorizes existing research by scope, methodological approach, and evaluation maturity. The review concludes by identifying key research directions—such as lightweight consensus protocols, energy-efficient optimization, and cross-chain interoperability—to advance the development of practical, secure, and privacy-preserving SL systems.
Browser extensions are commonly used to improve the browsing experience and accessibility. However, installing extensions naturally increases the user’s risk. This work presents AXECC , a novel framework for measuring the impact of web tracking and accessibility. The AXECC framework includes static, dynamic, and accessibility analyses across multiple web pages. We utilise the AXECC framework to analyse the web tracking and accessibility impact in the wild on 21k real-world extensions collected from the Chrome Web Store. In our analysis, we identify that 15.97% of extensions (with more than 600M users) perform a type of third–party tracking in the first 60 seconds after loading. These results are highly correlated with the extension category and are consistent across different web pages. Furthermore, we find that a small number of extensions (with 65M users) alter the accessibility of a web page when browsing, and these alterations are often complex and involve more tracking. Our work reveals a significant privacy risk from web tracking within popular browser extensions, often intertwined with complex accessibility alterations.
End-to-end machine learning malware detectors are vulnerable to adversarial EXEmples, carefully-crafted malicious programs that evade detection through minimal perturbations. Such attacks typically operate by either replacing unused content (patch attacks), or injecting new patterns (content-injection attacks). To counter these attacks, recent research has focused on certification methods for end-to-end models that aim to prove robustness guarantees within bounded perturbation sizes. However, existing approaches (i) are not robust against content-injection manipulations and (ii) provide only probabilistic guarantees for perturbations that are negligible relative to the overall program size. Hence, in this paper we address these limitations through a novel deterministic certification schema based on (de)randomized smoothing. Our defense splits each executable into non-overlapping chunks and classifies them independently. The final decision is obtained via majority voting across all chunks, ensuring that localized modifications, such as injected or patched code, influence only a limited subset of chunks and have minimal impact in the overall classification. This design guarantees that each chunk either contains or does not contain an adversarial perturbation, enabling us to (i) handle manipulations occurring at arbitrary locations within the program, and (ii) compute deterministic estimates of the perturbation magnitude required to evade detection. We demonstrate the effectiveness of our certification schema through extensive experimental analysis, comparing our defense against a range of state-of-the-art attacks and defenses. The results show that our approach achieves unmatched robustness across all tested attack scenarios, substantially outperforming competing defenses.
Malware authors increasingly exploit API Hashing to create ”invisible” system calls, replacing explicit function names with dynamically computed hashes that evade detection systems. This sophisticated obfuscation technique poses three critical challenges: accurately identifying hash functions within obfuscated code, linking computed hashes to their corresponding API calls, and detecting the growing diversity of hash algorithm variants. Existing rule-based approaches fail against these adaptive threats and cannot identify modern hash variants. We propose GAEDM , a novel framework that combines deep learning with program analysis to address these challenges. Our key innovation integrates static taint analysis with a genetic algorithm-enhanced assembly language model that generates diverse training variants, enabling robust detection of previously unseen obfuscation patterns. Experimental evaluation demonstrates that GAEDM achieves 91.9% MRR and 94.6% Recall@k in hash function identification, representing improvements of 18.4% and 8.2% respectively over state-of-the-art methods. GAEDM detects sophisticated obfuscation patterns that completely evade existing approaches, enabling security analysts to uncover previously undetectable threats and significantly advancing malware defense capabilities.
In this paper, we focus on the robustness of behavior-based malware analysis models, justified by the need to address the high mutation rates of malware executables that debilitate conventional signature-based approaches and even behavior-based AI solutions. In response to these challenges, we propose MAMBA + , an obfuscation-resistant dynamic analysis approach tailored for uncovering malware behavior. We have assembled a comprehensive collection of behavioral obfuscation attacks designed to undermine behavior-based models. The central concept behind MAMBA + involves treating obfuscated calls as perturbed data and introducing a novel loss function to effectively balance ground-truth predictions and the handling of these perturbations. To facilitate this approach, MAMBA + designs adapted embedding mechanisms to transform traces of API calls into high-dimensional vectors for attention calculations. Through a comprehensive empirical study with seven obfuscations and three unseen attacks, we reveal important qualitative properties of MAMBA + , and quantitatively demonstrate its superiority in performance and robustness to all compared methods.
The correct use of cryptography is central to ensuring data security in modern software systems. Hence, several academic and commercial static analysis tools have been developed for detecting and mitigating crypto-API misuse. While developers are optimistically adopting these crypto-API misuse detectors (or crypto-detectors) in their software development cycles, this momentum must be accompanied by a rigorous understanding of their effectiveness at finding crypto-API misuse in practice . This paper describes the MASC framework, which enables a systematic and data-driven evaluation of crypto-detectors using mutation testing. We ground MASC in a comprehensive view of the problem space by developing a data-driven taxonomy of existing crypto-API misuse, containing 107 misuse cases organized among nine semantic clusters. We develop 19 generalizable usage-based mutation operators and three mutation scopes that can expressively instantiate thousands of compilable variants of the misuse cases for thoroughly evaluating crypto-detectors. Using MASC, in a previous study, we evaluated nine major crypto-detectors and discovered 19 unique, undocumented flaws that severely impact the ability of crypto-detectors to discover misuses in practice. This paper substantially extends our MASC framework and offers updated evaluation of the crypto-detectors in our 2022 study, in addition to 5 more, major crypto-detectors. Through this work, we find 6 new, undocumented flaws, and demonstrate that these flaws affect the crypto-detectors regardless of their origin; open-source community, industry, and/or research. We conclude with a discussion on the diverse perspectives that influence the design of crypto-detectors and future directions towards building security-focused crypto-detectors by design.
The Android Permissions System (APS) is a permission-based access control mechanism that controls the applications’ access to protected resources such as user contacts and GPS locations. The evolution of APS -as a critical component of Android- has made the user experience more convenient. However, the vulnerabilities and attacks targeting the APS indicate that its thorough security analysis is essential to ensure system security and to protect user information. Although previous works applied formal methods to verify the security of APS, its new capabilities, such as One-Time Permissions (OTPs), have not yet been studied. In this paper, we present a model checking approach for security verification of APS that supports OTPs. We use Large Language Models in a Chain-of-Thought process to assist in extracting the behavior of APS from its documentation and source code to design a formal model and a set of security properties. Then we use the TLC model checker to verify the security properties against the model. We analyze the APS in both Android 6 and Android 11 and we succeed in discovering a previously known vulnerability in Android 6 as well as a new vulnerability in One-Time Permissions in Android 11 named Permanent Permission Access (PPA). We also implement an exploit code to show that PPA leads to permanent illegal access to sensitive resources. Our experiments show that the developed exploit app works on the latest versions of Android too (including Android 15). Besides, we propose the required fix to mitigate the vulnerability.
