157082 research outputs found
Sort by
Generating Traffic-Level Adversarial Examples from Feature-Level Specifications
No full textMachine learning-based network intrusion detection methods often rely on statistical summaries of traffic, causing a disconnect between the traffic space and the feature space that is difficult to bridge [13]. Realistic adversarial attacks are hard to generate because natural well-formedness constraints at the traffic level aren’t respected at the feature level with usual adversarial attack generation methods. We use a novel attack generation method combining two tools: (1) a bespoke synthetic traffic generation suite, PackGen, and (2) a formal verification tool for neural networks, Vehicle [7]. PackGen produces aggregated Markov chain representations of network traffic which allows us to reconstruct valid packet sequences that are modified by realistic perturbations on an input specification. Vehicle’s formal specification language lets us represent granular threat models such as adversaries who can only manipulate packet timings. Unlike other methods, Vehicle’s formal verification is guaranteed to find counterexamples if they exist, which correspond with evasive adversarial examples. We feed these feature-level counterexamples into modified PackGen representations to generate PCAP files containing reconstructed, evasive network flows, generating adversarial examples that cross the gap between the traffic and feature spaces. We evaluate PackGen by replicating DoS traffic using a variety of timing distributions, before testing our full pipeline by producing evasive counterexamples, outperforming projected gradient descent.</p
Generating Traffic-Level Adversarial Examples from Feature-Level Specifications
No full textMachine learning-based network intrusion detection methods often rely on statistical summaries of traffic, causing a disconnect between the traffic space and the feature space that is difficult to bridge [13]. Realistic adversarial attacks are hard to generate because natural well-formedness constraints at the traffic level aren’t respected at the feature level with usual adversarial attack generation methods. We use a novel attack generation method combining two tools: (1) a bespoke synthetic traffic generation suite, PackGen, and (2) a formal verification tool for neural networks, Vehicle [7]. PackGen produces aggregated Markov chain representations of network traffic which allows us to reconstruct valid packet sequences that are modified by realistic perturbations on an input specification. Vehicle’s formal specification language lets us represent granular threat models such as adversaries who can only manipulate packet timings. Unlike other methods, Vehicle’s formal verification is guaranteed to find counterexamples if they exist, which correspond with evasive adversarial examples. We feed these feature-level counterexamples into modified PackGen representations to generate PCAP files containing reconstructed, evasive network flows, generating adversarial examples that cross the gap between the traffic and feature spaces. We evaluate PackGen by replicating DoS traffic using a variety of timing distributions, before testing our full pipeline by producing evasive counterexamples, outperforming projected gradient descent.</p
Detection of Glaucoma Using OCT Images
No full textGlaucoma is a chronic eye condition that results in damage to the optic nerve, which leads to permanent blindness if it is not diagnosed and treated early. As of now, there are no treatments for curing glaucoma; thus, early detection might halt its progression. There are several imaging techniques used in detection of glaucoma, and optical coherence tomography (OCT) has gained prominence due to its ability to provide high-resolution images. However, manual observation and diagnosis by ophthalmologists are labor-intensive. This study proposes a novel method for automated detection of glaucoma by using YOLOv5s, which is an object detection model, applied to OCT images of patients both with and without glaucoma. The preliminary results are encouraging and achieved a detection accuracy of 99.3%
TS-Net: An Emotion Recognition Network Based on Temporal-Spatial Features of EEG Signals
No full textEEG-based emotion recognition can effectively monitor the users’ real-time emotional states, provide more objective physiological data support for mental health assessment, and thus detect the users’ potential psychological problems in a timely manner. Although existing research has made notable advancements, The biological and topological information among brain areas has not been sufficiently utilized. To the end, the paper proposes an emotion recognition network, dubbed TS-Net, based on the temporal-spatial features of EEG signals. TS-Net contains a special-purpose temporal feature extraction component (1DCNN) and a special-purpose spatial feature extraction component (gMLP), which enable it to fully analyze users’ emotional states based on the neural activity intensity in different brain functional areas. Experiment results show that TS-Net reached an overall accuracy of 97.87%, 96.79%, and 97.99% for arousal, valence, and dominance evaluated with the dataset DREAMER, respectively, which demonstrates that TS-Net has outperformed the existing advanced methods for emotion recognition. Finally, we conducted tests on two self-collected emotion classification datasets, and our model also achieved satisfactory results
Persistence of Links in Risk-Sharing Networks: Evidence From Rural Ethiopia
No full textMost rural households in developing countries rely on agriculture for their income. They face related risks due to various shocks. In rural Ethiopia a majority of surveyed households reported to have experienced rainfall related shocks, about a third reported crop pest or diseases related shocks. Given that asset, insurance and credit markets are either missing or poorly developed, households face significant constraints in insuring themselves against these shocks. They often rely on the accumulation and depletion of assets (livestock, etc.), savings, borrowing, diversification of economic activities and on risk-sharing networks. Using a unique feature of an Ethiopian longitudinal dataset collected in 2004 and 2009, we investigate the persistence of links within households' risk-sharing networks. We do this by looking at two types of attributes: i) household characteristics such as demographics, assets, location; ii) link attributes such as relationship between households, types of arrangement (money lending, labor-sharing, etc), co-membership in local groups as well as observable differences in income and assets holdings. We investigate whether households sustain these links more on the basis of economic or financial factors such as wealth or more on the basis of social factors such as geographical proximity and shared kinship. Using logit estimation techniques, we find that many of our proxies for social factors play a significant role in the persistence of links in risk-sharing networks. Livestock and land endowments do not seem to play an important role. Local or governmental institutions aiming at bolstering persistence of informal risk-sharing arrangements could focus on networks which have a more local geographically dense structure, or one based on shared kinship. To our knowledge, this is the first time that both the 2004 and 2009 rounds of this large survey are combined. They have a rare feature: they are refined enough to allow us to identify precisely links or the individuals on whom one relies in case of needs. We can thus identify precisely which links persist (being reported both in 2004 and 2009) and which attributes significantly impact persistence. Given this particular attribute of this dataset we have yet to see in the literature a comparable analysis
Exploring novel thiazole-based minor groove binding agents as potential therapeutic agents against pathogenic Acanthamoeba castellanii
No full textDue to limited advances in diagnosis and targeted therapy, as well as poor understanding of pathophysiology, infections due to Acanthamoeba have remained a medical concern. With their ability to selectively bind to DNA sequences, minor groove binders have emerged as useful therapeutic agents against parasitic infections. Herein, 6 novel thiazole-based minor groove binders were synthesized. Purification of intermediate compounds was accomplished by utilising silica gel column chromatography, while thin-layer chromatography was utilised to monitor reactions. The purification of the final products was achieved using liquid chromatography. Confirmation of structures was achieved by NMR spectroscopy and mass spectrometry. All compounds were evaluated against pathogenic A. castellanii via in vitro assays. At micromolar concentrations, selected minor groove binder derivatives revealed potent effects against (i) A. castellanii trophozoites as observed using amoebicidal assays, (ii), against A. castellanii cysts as observed using excystation assays, and (iii) against A. castellanii-mediated host cell death utilising human cerebrovascular endothelial cells, but (iv) showed limited effects against host cells alone, using cytotoxicity assays. The binding interaction between minor groove binders and DNA was studied using isothermal titration calorimetry and molecular docking simulations to provide insights into their binding affinity and mode of interaction. The findings of our study underscore the therapeutic value of thiazole-based minor groove binders as potent agents against A. castellanii, demonstrating effective antiamoebic activity with a low propensity for human cell damage, thus supporting their further development as antiamoebic agents.</p
An effective combination of mechanisms for particle swarm optimization-based ensemble strategy
No full textA high-quality ensemble strategy can effectively integrate several coefficients, mechanisms, and algorithms into a single framework. The adaptability, timing of intervention, and complementarity are the key factors to consider for the selected coefficients, mechanisms, and algorithms. In this study, two complementary variants based on Particle Swarm Optimization (PSO), namely Modified PSO (MPSO) and Social Learning PSO (SLPSO), were selected, forming IMPSO and ISLPSO after improvements. IMPSO excels at exploration, while ISLPSO excels at exploitation. The Improved Novel Ratio Adaptation Scheme (INRAS) is employed as a selection strategy and provides the ability to abandon less-optimal particles. The Modified Nonlinear Population Size Reduction (MNLPSR) enables the extension of generations, allowing for more sufficient evolution in later stages. Due to the use of MNLPSR, an improved inertia weight and adaptive acceleration coefficients are introduced to ensure compatibility with the proposed algorithm. Additionally, an improved dynamic differential mutation strategy is designed not only to be compatible with the proposed algorithm but also to enhance particle diversity. Both the Improved Sine Cosine Algorithm (ISCA) and Sequential Quadratic Programming (SQP), which focus on searching near the global best particles, are incorporated into the proposed ensemble strategy. This PSO-based variant is named the Effective Combination of Mechanisms for a PSO-based Ensemble Strategy (ECM-PSOES). Ablation experiments demonstrated the effectiveness of the individual coefficients and mechanisms. The novel PSO-based variant was evaluated on the CEC2017 benchmarks and compared with 14 state-of-the-art PSO-based variants and 11 non-PSO algorithms. Additionally, to evaluate the flexible and robust capability of the proposed algorithm, three real-world applications for long-term Transmission Network Expansion Planning (TNEP), Planetary Gear Train Design (PGTD), and Robot Gripper Design (RGD) were tested. The experimental results illustrate that the proposed algorithm displays superior performance compared to recently proposed PSO-based variants and most non-PSO algorithms. However, the proposed algorithm falls short of outperforming Differential Evolution (DE)-based algorithms and still requires time to match the performance of top-tier metaheuristics. The source code of ECM-PSOES is provided at https://github.com/microhard1999/CODES
AV-SLAF:A Scenario-Layered Framework for Safety Analysis of Autonomous Vehicles Based on STPA and CTA
No full textEnsuring safety in autonomous vehicles (AVs) requires addressing hazards beyond functional failures, especially those arising from Performance Limitations (PLs) and Triggering Conditions (TCs) under varying Operational Design Domains (ODDs). This paper proposes AV-SLAF, a scenario-layered safety analysis framework integrating System-Theoretic Process Analysis (STPA) with Cause Tree Analysis (CTA) and internal algorithm modeling. By incorporating layered ODD scenarios into the control structure and modeling internal logic of AV modules, AV-SLAF systematically identifies PLs and TCs critical for Safety of the Intended Functionality (SOTIF). Unlike traditional methods focusing solely on structural-level interactions, the proposed framework bridges external scenario modeling with internal algorithms, enabling a more complete view of hazard propagation. A case study on autonomous port vehicles demonstrates the framework’s applicability, yielding a structured set of 84 PL-TC pairs and a partial cause tree for the Planning and Control module. The resulting causal structure reveals dependencies among algorithmic components and their safety-relevant conditions. The proposed framework enhances the traceability and completeness of safety analysis for complex AV applications.</p
Amplification of Nonlinear Response of Floating Photovoltaics by Coastal Topography:Experimental and Numerical Study
No full textNearshore coastal regions have become popular for floating photovoltaics (FPV) installations. During propagation over seabed topography towards nearshore FPV systems, waves undergo intricate transformations by shoaling, reflection and refraction, potentially influencing hydrodynamic responses of these emerging marine renewable energy structures in ways that are not well understood. Therefore, wave flume experiments and multiscale fully coupled time-domain fluid-structure interaction (FSI) simulations are performed to examine the topography effect on the nonlinear responses of nearshore FPV systems at a field site in the East China Sea. Experimental results reveal that near-resonant wave interactions in coastal regions drive significant energy transfer among different wave frequencies, amplifying the nonlinear dynamic responses of FPV systems by channeling energy toward their natural modes. As a result, second-order heave and pitch responses are amplified by up to 117.87% and 136.38% compared to the case without topography, which in turn lead to an increase in mooring tension. Moreover, the topography-induced amplification of nonlinear wave harmonics enhances the surge mean drift of FPV. This enhancement exhibits a negative correlation with the relative FPV length with respect to the wavelength. Comparisons between experiments and fully coupled simulations for irregular waves indicate that neglecting topography causes the FPV dynamic response model to produce inaccurate estimations of heave/pitch motions, while FSI simulations forced by high-fidelity local wave fields predicted by the fully nonlinear Boussinesq wave model are capable of capturing the observed topographic effect. These findings provide the theoretical basis for design consideration of the safe, cost-effective deployment of efficient FPV systems in coastal waters
The missing puzzle piece: contextual insights for enhanced pharmaceutical supply chain forecasting
No full textAccurate forecasting in pharmaceutical supply chains is essential for ensuring medicine availability, particularly in low-resource settings. However, many existing approaches rely solely on historical consumption data, provide only point forecasts and fail to account for the operational context and uncertainty inherent in these environments. In this study, we collaborated with experts at the Ethiopian Pharmaceutical Supply Service to identify key contextual factors, such as stock replenishment cycles, fiscal inventory counts and seasonal disease outbreaks and integrated them into forecasting models. Using five years of monthly distribution data (December 2017 to July 2022) for 33 essential medicines, we evaluated a range of forecasting methods, including statistical, machine learning and foundational models. We assessed point and probabilistic forecast accuracy using standard evaluation metrics. Our findings show that incorporating contextual variables significantly improves forecast performance, especially for classical time series models. We recommend investing in the routine collection of contextual indicators and adopting transparent, low complexity forecasting methods that can be sustained in practice. To support reproducibility and wider use, we provide all data, code and the full manuscript as an open, executable Quarto project developed in R and Python
Access Repository Dashboard
Do you manage Heriot Watt Pure? Access insider analytics, issue reports and manage access to outputs from your repository in the CORE Repository Dashboard!