Homomorphic encryption (HE) enables secure computation on encrypted data, making privacy-preserving machine learning (ML) viable for sensitive applications. Non-linear activation functions (AFs) such as ReLU, Sigmoid, and Tanh remain costly, as traditional polynomial approximations incur deep circuits and long runtimes. We propose a GPU-accelerated method that approximates these activations using fast piecewise-linear functions, formulated as a sign-decision problem to simplify computation and reduce circuit complexity. Compared to existing methods, our technique achieves a lower mean squared error (MSE) to the true activation functions, reducing MSE by approximately 55% with an error below 10(-2). Experiments demonstrate speedups of up to 128x on MNIST and 840x on CIFAR-10, with accuracy comparable to state-of-the-art methods. Across various neural architectures and datasets, our pipeline significantly reduces homomorphic evaluation time and circuit depth compared to polynomial approximations of similar accuracy.
Black-box Artificial Intelligence (AI) systems have achieved state-of-the-art accuracy in many problem domains in recent years. However, the lack of transparency of these systems is a bottleneck in their usage in high-risk applications which make automated decisions on individuals. Trustworthy AI proposes principles such as reliability, validity, privacy, fairness, and explainability among others, to mitigate risks from large-scale AI deployments in such domains. Explainable AI (XAI) is a technique of providing insights into the decision-making process of black-box systems thus enabling transparency. It plays a crucial role in communicating the rationale of automated decisions to relevant stakeholders. Though explainability is a highly desirable requirement, recent research has determined that explanations can introduce new privacy risks in AI systems. Researchers have demonstrated different types of privacy attacks on XAI deployed in production and cloud systems. Despite these risks, currently there is a lack of research into defenses for known privacy attacks in XAI. In this article, we contribute to this gap by proposing a defense mechanism for attribute inference attack on feature-based XAI. We empirically evaluate a well-known privacy preservation technique, namely, additive noise, and show its impact on privacy, explainability and utility. Our findings indicate that additive noise enables privacy while achieving faithful explanations and without compromising model utility.
User profiling based on browsing behavior has traditionally been applied to improve web personalization and marketing strategies. However, leveraging browsing patterns to assess cybersecurity risks remains underexplored. In this paper, we propose a profiling framework based on domain name system (DNS) traffic analysis. Our approach models user browsing behavior using two main factors: browsing intent and domain reputation. By aggregating risk weights derived from accessed domains, we compute a personalized browsing risk score that reflects the user’s exposure to online threats. We validate the effectiveness of our framework through experiments that demonstrate its ability to differentiate users with varying levels of browsing risk. Our findings offer new insights into user-centric cybersecurity assessment using minimal yet meaningful data sources.
In an era of increasing digital privacy risks, images shared online can inadvertently reveal sensitive location data through identifiable elements such as logos, road signs, and text. These disclosures enable unauthorized tracking and data mining, raising serious privacy concerns. This paper proposes a novel framework that combines object detection and generative inpainting for privacy preserving image reconstruction. A Mask R-CNN model is trained on three diverse datasets to detect and segment location identifiable elements accurately. The detected regions are then de-identified using a proposed Denoising Diffusion Probabilistic Model (DDPM)-based inpainting method, which preserves scene integrity by ensuring geometric consistency and natural lighting. Unlike traditional inpainting methods, the proposed framework dynamically refines image reconstructions through controlled denoising, achieving high realism. The effectiveness of the method is evaluated using standard image quality metrics, including PSNR, SSIM, and FID, alongside subjective visual assessments. Experimental results show that the proposed approach outperforms baseline models such as CNNs and GANs, offering a robust solution for privacy preserving image reconstruction.
Online review platforms have drastically reshaped how we interact, make purchasing decisions, and engage with digital content. However, the rise of deceptive content, privacy breaches, and misinformation has undermined online content credibility, impacting the trustworthiness of the information shared. To address these issues, we propose a robust framework for automated trust assessment of online reviews, focusing on identifying deceptive online content. The core of our approach is the trust vector, a novel feature representation that captures key user engagement factors influencing content trustworthiness. By applying the Weighted Trust Scoring Method (WTSM), we calculate a weighted trust score that strengthens the model's interpretability and effectiveness in trust evaluation. The proposed model is evaluated on two benchmark datasets-the Deceptive Opinion Corpus Dataset and the Yelp Review Dataset. The framework achieves classification accuracies of 85% and 87%, respectively, demonstrating its effectiveness in distinguishing deceptive from truthful content.
Existing fact-checking datasets suffer from limitations, such as lack of metadata annotations or insufficient evidence. In this paper, we introduce a comprehensive dataset comprising approximately 5,000 real-world claims, collected from Politifact and Snopes, with the corresponding evidence scraped for each claim. Our dataset uniquely includes detailed metadata such as source credibility metrics (domain age, top level domain score, page rank, bias rating score and factual rating score). Additionally, each claim is segmented into standalone statements, generated using a Large Language Model in order to facilitate efficient evidence retrieval. We further provide an impact analysis score that represents the potential societal influence of each claim. We also develop a baseline automated fact checking pipeline using the proposed dataset.
Anomaly detection in Open Charge Point Protocol (OCPP) is key to securing the infrastructure of electric vehicle charging stations (EVCS), which are increasingly vulnerable to cyber threats. Anomaly detection enables cyber-attack detection, and supervised learning is commonly used. However, its dependence on labeled datasets, often sparse and unbalanced, limits its detection ability on advanced threats. This study highlights the limitations of supervised learning techniques and explores clustering in semi-supervised and unsupervised learning techniques. To our knowledge, no previous work focuses on anomaly detection using clusters for semi-supervised and unsupervised learning applied to the CICEVSE2024 dataset. Our key contribution enhances OCPP security, enabling more robust anomaly detection in EVCS infrastructure and the broader smart grid ecosystem.
Given the escalating global threat of phishing emails, it is imperative to develop effective solutions to mitigate their potentially devastating impacts on society. This study endeavours to construct a federated multilingual spam detection system employing logistic regression, specifically targeting English, French, and Russian emails. This is the first work to the best of our knowledge which considers a non-deep learning setting for federated learning, and combines federated learning with multilingual phishing detection. Evaluation of the models is based on accuracy metrics which are compared with a most frequent class baseline. Our findings indicate that an optimal configuration comprises 10 clients undergoing 100 epochs of training with 100 rounds of federated learning, resulting in superior performance. Notably, this approach significantly outperforms the baseline, achieving an accuracy of 89.46% compared to 70%.
Personal Identification Numbers (PINs) are widely used for authentication on mobile devices such as smartphones, which act as gateways to many important accounts (e.g., financial, email, etc.). Unfortunately, people tend to choose easy-to-recall PINs involving birthdays, anniversaries, or keypad patterns that are vulnerable to guessing attacks. System-assigned PINs can improve PIN security in this regard; however, they have usability problems such as feeling the need to store the assigned PIN, longer login times, and difficulty remembering. In this paper, we propose, design, and evaluate a set of short training techniques (16-34 seconds) inspired by implicit learning techniques, to improve the usability of system-assigned PINs. We evaluated our designs in a two-session user study with 184 university students. Our results show that some designs offer significant improvements in the login success rate, login times, and user perceptions. These advantages are in addition to our design’s short single-session training, making it more compatible with typical registration workflows than previously proposed multisession training techniques.
As cyber threats continue to evolve, there is a need for autonomous cyber defense (ACD) strategies capable of fast and context-aware responses. Reinforcement learning (RL) has shown promise for automating cyber defense by exploring and learning effective countermeasures, yet it often struggles with sparse reward signals and insufficient context to handle diverse attack scenarios. Furthermore, the convergence time taken by an RL agent is often high, which makes it difficult to train the RL agent in online settings. To address these challenges, we propose a large language model (LLM)-enhanced RL method that builds and queries a knowledge graph (KG) derived from agent-environment interactions. We leverage the pre-trained knowledge of an LLM on different cybersecurity frameworks and use the LLM to analyze a part of the KG to generate appropriate actions for the RL agent. We infuse the knowledge extracted from the LLM into the RL agent's training loop in two ways. First, the state vector of the RL agent is augmented with the most effective action and its corresponding reward, as determined from the KG. Second, the suggested action from the LLM is used as a reference policy. In addition, we introduce a regularization term in the loss function to make the RL policy close to the reference policy. To validate our approach, we develop a custom RL environment guided by the MITRE ATT&CK framework, enabling the agent to generate tailored mitigation strategies for detected cyber attacks. Experimental results show that our proposed approach significantly outperforms the baseline RL by over 75% in terms of taking better mitigation actions.
This paper proposes a dynamic, decentralized social trust model to meet Know Your Customer (KYC) and Anti-Money Laundering (AML) compliance requirements, especially for individuals lacking formal documentation. Building on previous work, we present a conceptual model for dynamically computing trust and risk values to support financial transaction approvals through innovative recognition schemes. Agent-based simulations across four scenarios demonstrate that dynamic trust mechanisms enhance transaction approval rates and payment volumes, provided malicious agents are controlled. The model is flexible, adaptable to diverse populations, and promotes financial inclusion without compromising compliance requirements. It also offers a foundation for future research on mitigating risks from malicious actors. Furthermore, it is particularly suited for enabling compliant and inclusive decentralized finance (DeFi) applications.
The development of industrial control systems (ICS) has led to security vulnerabilities in ICS protocols, which have caused significant threats to these systems. Fuzzing is a highly effective technique for detecting vulnerabilities by exposing the target program to a large volume of malformed and unexpected input. Traditional fuzzing methods are time-consuming and have low acceptance rates and coverage due to the manual extraction of specifications, lack of guidance, and blind mutations, limiting their efficiency in discovering vulnerabilities. Currently, the Generative Adversarial Network (GAN)-based fuzzing methods are used to learn the syntax and format of the input data and generate valid test cases. However, they still suffer from low diversity in the generated test cases, which can affect the potential of triggering vulnerabilities. We have proposed a transformer-based fuzzing framework called TFVDFuzzer to solve these issues. In this framework, we leverage the Transformer model to automate the learning of ICS protocol specifications, especially Modbus, significantly speeding up the fuzzing process and enhancing the test case acceptance rate. Furthermore, we used a deep learning model that employs an attention mechanism to determine the appropriate bytes for mutations to enhance the diversity of the test cases. The TFVDFuzzer is assessed using a benchmark Modbus dataset. It outperforms Peach, Fuzzowski, and NCMFuzzer with significant improvements, achieving a Test Case Reception Rate (TCRR) of 98.15% compared to their highest values of 51.42%, 69.02%, and 94.25%, respectively, and a Test System Anomaly Rate (TSAR) of 3.1% relative to their highest values of 0.445%, 0.475%, and 0.96%, respectively.
Insider threats, where legitimate users misuse their access for malicious purposes, remain challenging to detect due to their contextual and behavioral subtleties. This paper presents a novel machine learning framework that captures user activity sequences through a user-centric representation named the User Daily Activity Sentence (UDAS). Unlike prior work that informally uses daily sequences, we formalize UDAS as a behavioral encoding technique using Word2Vec embeddings and extensively evaluate it across multiple unsupervised anomaly detection methods.To enrich this representation with relational context, we propose a graph-based extension that constructs a user interaction graph based on co-device usage and domain access. A Graph Convolutional Network (GCN) is applied to enhance semantic user embeddings, and anomaly detection is performed using Kmeans clustering.To the best of our knowledge, this is the first work to systematically combine semantic sequence embeddings with graph-based relational learning for insider threat detection. Experiments on the CERT Insider Threat v4.2 dataset show that our method outperforms prior unsupervised models in accuracy and robustness. The proposed framework requires no feature engineering or labeled data, making it applicable to real-world monitoring environments.
The integration of Information and Communication Technology (ICT) into Smart Grids has revolutionized the efficiency and functionality of power systems. However, this advancement has also introduced significant cybersecurity challenges. Among the most vulnerable components are digital substations, which serve as critical hubs in power distribution and are susceptible to cyberattacks that can trigger cascading failures and widespread disruptions. This paper presents a structured procedure for risk assessment, emphasizing system analysis, dependency evaluation, and asset profiling based on their vulnerability to potential adversarial techniques. Prioritizing risks facilitates targeted and effective post-attack mitigation strategies, ensuring faster recovery, reduced impact, and minimized downtime following an attack. Using the MITRE ICS ATT&CK Framework, we systematically identify adversarial techniques and assess the criticality of substation components. Different substation devices attract specific attack techniques depending on their role and exposure in the system; the MITRE ICS framework helps map these patterns to enable focused and effective defense. With a clear understanding of system structure and threats, we can develop mitigation solutions tailored to specific needs. RF-RADS generates a quantified risk profile by scoring substation assets, enabling systematic identification of critical components, with control servers, workstations, and data gateways identified as the highest-risk assets.
As quantum computing technology continues to advance towards a cryptographically relevant scale, cybersecurity for critical infrastructure such as electrical power systems must prepare for an existential threat. A crucial step to mitigating the potential damage a quantum computer-aided attack may cause is identifying quantum-vulnerable algorithms that are currently used in the standards for the security of power system control centres and communication networks. The IEC 62351 is a series of standards for protecting data in power automation systems and includes several quantum-vulnerable cryptographic primitives. Parts 7, 8, and 9 of the IEC 62351 are dedicated to network and system management, role-based access control, and key management. In this work, we conduct a comprehensive vulnerability assessment of the cryptographic algorithms selected in these parts of IEC 62351.
Visual speaker authentication (VSA) remains vulnerable to sophisticated spoofing methods, such as deepfakes. Traditional deep learning approaches require extensive userspecific enrollment data and show poor generalization to new speakers. In response, we employ a few-shot meta-learning technique, specifically Model-Agnostic Meta-Learning (MAML), integrated with dynamic lip movement analysis utilizing optical flow to develop a scalable anti-spoof VSA framework. We validated our model using the GRID audiovisual dataset, with spoofing attacks simulated via Wav2Lip for deepfake lip synchronization. The results demonstrate the model’s superior performance, evidenced by near-perfect classification accuracy and negligible error rates, addressing two key challenges of VSA: eliminating extensive training data while enabling rapid adaptation to unfamiliar speakers. Our approach significantly surpasses standard non-meta-learning approaches, substantiating its ability to address real-world scenarios.
Large Language Models (LLMs) have become central to modern NLP applications, yet their reliance on sensitive training data introduces significant privacy risks. Traditional approaches to differential privacy, such as DP-SGD, apply uniform noise at the gradient level and treat all features equally, ignoring the inherent correlations in structured sequence data. In this work, we propose a novel forward-pass privacy mechanism that applies per-attribute differential privacy across correlated sequence inputs. Our framework leverages the Matrix-Variate Gaussian (MVG) mechanism to inject structured, directional noise during the forward computation, enabling fine-grained privacy control that aligns with attribute sensitivity. Privacy budgets are assigned on a per-token basis using an Inverse Gaussian Distribution, allowing position-aware adaptation across input sequences. To propagate these budgets through the model, we introduce a data-independent Layer-Wise Contribution Propagation (LCP) algorithm that maps input sensitivity to output features, even in transformer architectures. We validate our method on the SST-2 sentiment classification benchmark, demonstrating improved utility over existing approaches such as DPSGD and DP-Forward, particularly under strict privacy regimes. Our results highlight the benefits of structured noise in preserving utility while ensuring strong, attribute-level privacy in models trained on correlated sequential data.
This research presents the preliminary findings of an ongoing project focused on the security of General-Purpose AI (GPAI) applications. We introduce three key contributions: (i) a taxonomy of GPAI-specific vulnerabilities, offering a structured classification of security risks unique to GPAI models and applications; (ii) a generalized GPAI application architecture, serving as a meta-model for analyzing a wide range of real-world use cases; and (iii) G-STAR, a novel threat modeling reference framework that identifies key entities and their interrelationships in GPAI ecosystems, and provides a structured methodology for assessing and mitigating potential threats. Our study addresses both data and model vulnerabilities inherent in GPAI systems, highlighting critical security challenges. While the research is still in its early stages, the initial results provide a valuable foundation for continued investigation. Future work will focus on enhancing the generalized architecture, exploring mitigation strategies in depth, and applying and refining the G-STAR framework in real-world GPAI scenarios. This work aims to support AI security practitioners in promoting secure development and deployment of GPAI systems across diverse domains.
The proliferation of Internet of Things (IoT) devices has been paralleled by a surge in sophisticated malware threats, posing significant challenges to traditional security mechanisms. Conventional malware classification models often depend on extensive labeled datasets and exhibit limited generalization capabilities, particularly when encountering novel or cross-architecture malware variants. In this study, we introduce TOMNet, a transductive meta-learning framework that synergistically integrates few-shot learning with graph-based inference to facilitate efficient IoT malware classification under datascarce conditions. TOM-Net employs a hierarchical GraphSAGE encoder to extract both structural and semantic features from function call graphs, augmented with adaptive similarity kernels for transductive label propagation and entropy-regularized decision boundaries to enhance open-set recognition. Empirical evaluations demonstrate that TOM-Net achieves a classification accuracy of 92.64% in the 5-way 10-shot setting under the closed-set condition, and an area under the curve (AUC) of 93.59% in the open-set setting, significantly outperforming state-of-the-art baselines in detecting previously unseen threats. These results underscore the practical applicability of TOM-Net for robust IoT malware defense in scenarios characterized by limited labeled data.
The use of synthetic data in health applications raises privacy concerns, yet the lack of open frameworks for privacy evaluations has slowed its adoption. A major challenge is the absence of accessible benchmark datasets for evaluating privacy risks, due to difficulties in acquiring sensitive data. To address this, we introduce SYNQP, an open framework for benchmarking privacy in synthetic data generation (SDG) using simulated sensitive data, ensuring that original data remains confidential. We also highlight the need for privacy metrics that fairly account for the probabilistic nature of machine learning models. As a demonstration, we use SYNQP to benchmark CT-GAN and propose a new identity disclosure risk metric that offers a more accurate estimation of privacy risks compared to existing approaches. Our work provides a critical tool for improving the transparency and reliability of privacy evaluations, enabling safer use of synthetic data in health-related applications. Our privacy assessments (Table II) reveal that DP consistently lowers both identity disclosure risk (SD-IDR) and membership-inference attack risk (SD-MIA), with all DP-augmented models staying below the 0.09 regulatory threshold.
