Image-based Completely Automated Public Turing Test to Tell Computers and Humans Apart (CAPTCHA)systems are increasingly vulnerable to Adversarial Machine Learning (AML) attacks, where adversaries subtly perturb CAPTCHA images to deceive automated systems. This work proposes a novel defense mechanism that integrates Generative Adversarial Network with Convolutional Neural Network (GAN-CNN) to detect and mitigate adversarial examples. The GAN component enhances robustness by learning to generate adversarial samples during training, while the CNN is trained to classify both genuine and adversarial CAPTCHAs accurately.CAPTCHA images were obtained from the publicly available CAPTCHA Version 2 Images Dataset on Kaggle. Adversarial training, a critical step that improves the model’s resilience by introducing adversarial cases during learning, is then applied to this data. During the preprocessing phase, normalization and picture refinement techniques are used to reduce adversarial noise, blur, and distortions, guaranteeing cleaner inputs for the model.The system’s central component is a hybrid GAN-CNN architecture, in which the CNN functions as a strong feature extractor and classifier that can recognize even highly distorted images, while the GAN creates realistic adversarial CAPTCHA samples to test the classifier. First, a detection module determines if an input CAPTCHA is hostile or valid. If accepted, the image is sent to the classifier, which uses the CNN robust features to produce the final prediction. To ensure the reliability of the model and effectiveness against hostile attacks, standard metrics like accuracy, precision, recall, and F1-score are employed to evaluate the model’s general performance.
Geopolymer concrete (GPC) has emerged as a sustainable alternative to ordinary Portland cement due to its reduced carbon footprint and efficient utilization of industrial and waste-derived materials. However, predicting its compressive strength remains challenging because of complex nonlinear interactions among binder composition, alkaline activator chemistry, aggregate proportions, and curing conditions. This study proposes a Long Short-Term Memory (LSTM)-based deep learning model for accurate prediction of geopolymer concrete compressive strength. A comprehensive experimental dataset incorporating fly ash, eggshell powder replacement, SiO₂/Na₂O ratio, fine and coarse aggregate contents, reaction liquid dosage, and curing age was utilized. Exploratory data analysis, including histogram distributions, kernel density estimation, correlation matrix evaluation, and multivariate pair plots, revealed structured experimental variation and strong interdependency among mix parameters. The LSTM model was trained and validated using optimized hyperparameters, and its performance was evaluated through MSE, RMSE, and R² metrics. Training and validation loss curves demonstrated stable convergence without overfitting. The strong agreement between predicted and experimental values, along with normally distributed residuals centered around zero, confirms the robustness and generalization capability of the proposed framework. The study highlights the potential of deep learning techniques for intelligent, data-driven design and optimization of sustainable geopolymer concrete systems.
The increasing environmental impact of ordinary Portland cement production and natural aggregate depletion has accelerated the development of sustainable concrete alternatives. This study presents the development of self-compacting alkali-activated concrete (SCAAC) incorporating rice husk ash (RHA) as a supplementary binder and engineered fly ash based artificial coarse aggregate (NACA) produced from hardened fly ash as a partial replacement of natural coarse aggregate. A comprehensive experimental program was conducted to evaluate fresh properties, non-destructive characteristics, mechanical performance, durability behaviour, and environmental sustainability. RHA replacement up to 15% improved particle packing and reaction efficiency, while NACA replacement enhanced performance up to an optimum level of 70%. The optimum mix (F35R15N70) achieved a 28-day compressive strength of 66.03 MPa, splitting tensile strength of 6.02 MPa, flexural strength of 8.72 MPa, and impact energy of 19.8 kN-m, along with a high ultrasonic pulse velocity of 4776 m/s. Durability studies demonstrated reduced water absorption (2.55%), lower sorptivity, and improved resistance to acid and sulphate attack, with lower mass loss and higher residual strength compared to the control mix. Environmental assessment revealed a reduction in embodied energy (EE) from 4152 to 4055 MJ/m 3 and a decrease in global warming potential (GWP) from 620.2 to 618.1 kgCO 2 e/m 3 . Response Surface Methodology (RSM) was successfully modelled and optimized the combined effects of RHA and NACA, identifying an optimum mix that balances fresh performance, mechanical strength, durability, and environmental efficiency. The results confirm the feasibility of utilizing agricultural waste and engineered artificial aggregates to produce high-performance and sustainable SCAAC.
This study investigates the low-temperature magnetic behavior and dielectric relaxation mechanism in the compositionally designed spinel ferrite Ni0.33Co0.33Zn0.33Fe2O4 (NCZFO), synthesized using sol-gel auto-combustion techniques. The X-ray diffraction pattern with Rietveld refinement confirms the formation of a single-phase, polycrystalline inverse cubic spinel structure of NCZFO belonging to the space group Fd3(-)m (No. 227), where Ni2+/Co2+/Zn2+ ions occupy the octahedral sites, and Fe3+ ions occupy both octahedral and tetrahedral sites. The Raman-active phonon modes (A(1g), F-2g, and E-g) are detected in Raman spectra, which serve as a signature of the cubic spinel structure of NCZFO. The scanning electron microscopy images demonstrated large, sub-micron-sized, tightly connected grains with irregular lumps and uniform elemental presence and distributions with their stoichiometric atomic and weight percentages. Magnetic ordering, spin dynamics, and intergrain interactions were investigated using magnetic measurements by DC magnetization vs. temperature and magnetization vs. magnetic field loops. An irreversible and freezing/blocking temperature near 358 K and 215 K is revealed by zero-field-cooled (ZFC) and field-cooled (FC) curves and it suggest the presence of surface spin disorder, domain wall pinning and magnetic anisotropy influenced by Ni, Co and Zn ions. As the temperature drops, hysteresis loops at a specific temperature exhibit increasing coercivity and remanent magnetization, which is consistent with temperature-induced reversibility and strong anisotropy at low temperatures. The temperature-dependent dielectric constant (epsilon(r)) and loss (delta) are enhanced due to thermally activated charge carriers, and a weak relaxation plateau similar to 260 K was observed. Non-Debye type dielectric relaxation and the intrinsic permittivity (epsilon(r)) were observed, which is unaffected by conduction or interfacial effects at low temperature. The activation energy (E-a) was calculated to be similar to 0.41 eV for hopping of Fe2+/Fe3+ ions in NCZFO. This investigation provides new insights into the magnetic and dielectric relaxation governing mixed spinel ferrites in spintronic and cryogenic electronics applications.
Early treatment through brain tumor diagnosis via medical imaging is crucial, while deep learning approaches in the area suffer the most from issues like overfitting, lack of interpretability, and computational costs. To address these issues, the lightweight multi-modal autoencoded quantum dilated attention convolutional neural network (LMAQDAC), a lightweight model that enhances tumor detection by combining in a single efficient pipeline segmentation, feature extraction, and classification, is presented. The modified Hiking optimization algorithm (MHOA) further optimizes this method in terms of training speed and stability. Interpretability is improved by Gradient-weighted Class Activation Mapping (Grad-CAM), which highlights tumor regions, thus supporting clinician trust in the AI method. Model compression techniques are employed, enabling real-time performance, drastically reducing computation and memory overhead. This allows for deployment on small edge devices, thus providing localized, fast, and secure inference. The proposed method has 99.96
