Neural speech synthesis has increased the risk of synthetic spoken-command attacks on voice-driven systems, creating a need for robust anti-spoofing methods. This study proposes an external-aware Time–Frequency Short-Time Fourier Transform (TF-STFT) anti-spoofing framework that combines Generative Adversarial Network (GAN)-based TF-STFT synthesis with deep Spoken-Command Convolutional Neural Network (SC-CNN) operating on a three-channel spectro-temporal representation comprising TF-STFT magnitude, temporal-delta, and frequency-delta maps. By integrating static spectral information with dynamic spectro-temporal features, the proposed representation enhances discrimination between bona fide and synthetic speech. Experiments on a prepared subset of the Speech Commands dataset containing 20 command classes achieved 99.95
The growing prominence of data-driven artificial intelligence (AI) in healthcare has introduced significant challenges, including limited availability of high-quality training data, strict privacy regulations, and the complex nature of clinical information. This paper presents a data-driven hybrid generative framework for synthetic healthcare data generation that addresses these challenges through an integrated optimization strategy. The proposed framework combines diffusion-based stochastic generation and transformer-based autoregressive conditioning within a unified architecture, complemented by adversarial and latent-variable modeling to enhance data fidelity and structural consistency. The model effectively captures hierarchical, temporal, and contextual dependencies of heterogeneous healthcare data, including electronic health records (EHRs), medical imaging, and sequential patient trajectories. The architecture adopts a dual-pathway design, where the diffusion component models stochastic generative processes, while the transformer component learns structured conditional dependencies. Additional components, including Conditional GANs for label-aware synthesis and Variational Autoencoders (VAEs) for latent health state modeling, are incorporated in a modular manner to improve generative performance. To ensure privacy, differential privacy is integrated into the training process, while membership inference risk is analyzed and empirically constrained using Jensen–Shannon divergence-based measures. Extensive experiments conducted on real-world healthcare datasets demonstrate that the proposed framework consistently outperforms state-of-the-art generative models, including GAN, IGAMT, EHR-M-GAN, and EHR-Safe, across multiple evaluation criteria such as statistical fidelity, temporal consistency, predictive utility, and privacy preservation. The implications of these results point to the effectiveness of the framework in producing data that is clinically rule-following and statistically representative in nature, as evidenced by the application of quantitative metrics of fidelity and consistency of these data with clinical rules as well as the protection of privacy for all modalities being studied.
In digital systems, micropayments refer to extremely small transactions which occur frequently. These micropayments support a lot of existing systems such as pay-per-API billing, machine-to-machine interaction, and incentive mechanisms in distributed networks. Probabilistic micropayment systems enhance the system, by settling only a subset of transfers, although the rules for these probability are either prefixed or purely cryptographic. Due to this the system loses governance and limits auditability, and leaves with no explanation of the final win/lose outcome. This paper introduces CTIA (Cryptographic Ticket Issuer Agent), focused on a decision-layer component that assigns per-ticket win probabilities from telemetry while leaving the cryptographic ticket format and deterministic lottery verification unchanged. CTIA uses a Random Forest regressor to map telemetry (network load, relay reputation, budget utilization, and recent win rate) to a probability p, and logs SHAP attributions per issuance as an auditable explanation log. We evaluated CTIA under low-probability micropayment regimes under controlled simulations and compared it to static, analytic and alternative machine-learning controllers. With the results we found that CTIA preserves the intended probability and payout scale while providing adaptive and auditable probability assignments. However we noticed that fairness outcomes varied across regimes and were not consistently better than simpler controllers. During the experiment, we also observed that adaptive probability assignment can affect participant-level and temporal reward distribution, but the magnitude and direction of the effect were dependent on the probability regime, controller and evaluation horizon. Thus though this study we would like to position CTIA as an auditable probability-assignment layer.
Lung cancer is one of the leading causes of cancer-related deaths worldwide, mainly due to late diagnosis and limits of traditional screening methods. The traditional risk models based on demographics and smoking history often miss early behavioral and symptom-based signals. In this study, we propose an interpretable data-driven analytics framework for the lung cancer risk prediction by integrating heterogeneous behavioral, symptomatic, and clinical data to support informed decision-making. The proposed framework incorporates systematic data preprocessing, feature engineering through interaction terms (e.g., smoking–alcohol consumption), and data normalization to enhance predictive modeling. A dataset comprising 309 patient records was analyzed, ensuring data completeness and balanced representation of variables. Three machine learning models, namely, Random Forest, Logistic Regression, and CatBoost, were implemented and evaluated using the standard metrics such as precision, recall, F1-score, accuracy, and AUC-ROC. Among these, the CatBoost method achieved the highest performance, with an accuracy of 95.37
Modern vehicles increasingly rely on in-vehicle communication networks such as the Controller Area Network (CAN), which lack built-in security mechanisms and remain vulnerable to cyberattacks. This paper proposes a lightweight Intrusion Detection System (IDS), termed LR-ResNet-IDS, based on a reduced ResNet deep learning architecture for real-time CAN bus attack detection. The proposed framework includes a frame-building module that converts raw CAN traffic into structured grid representations, enabling automatic feature learning without manual feature engineering. The model is evaluated using a real-world CAN dataset collected from a physical vehicle under controlled attack scenarios. Experimental results demonstrate strong detection performance, achieving an accuracy of 0.955, precision of 0.948, and recall of 0.943, while maintaining low computational complexity. These results indicate that the proposed LR-ResNet-IDS provides an effective, lightweight, and scalable solution for real-time intrusion detection in automotive networks.
