Volume 14 Issue 3

A Multi-Branch Deepfake Detection Framework Using Spatial, Noise, And Frequency Domain Features With Attention Fusion

Authors: Lakshay Bhardwaj, Rishabh Jain, Kritika, Ritesh Kumar

Abstract: Deepfakes have emerged as a major challenge in digital media forensics because modern generative models can produce highly realistic fake facial content that is difficult to distinguish from authentic media. Their rapid growth has increased the risk of misinformation, identity misuse, and security threats in online environments, motivating the need for reliable forensic datasets and detection frameworks [1], [7], [10]. Existing deepfake detection methods often perform well only under controlled settings and show limited robustness when evaluated on unseen manipulations or datasets. Many CNN-based methods achieve high accuracy on known datasets but fail to generalize to unseen data. Prior works based on compact CNNs, frequency-aware learning, and multi-branch detection have shown promising performance, but cross-dataset generalization remains a major challenge [2]–[5], [8]. To address this issue, this work proposes a multi-branch deepfake detection framework that jointly learns from spatial appearance information, residual noise traces, and frequency-domain artifacts. The spatial branch uses a pretrained EfficientNet-B0 backbone to capture facial inconsistencies [6], the noise branch extracts forensic residual cues using SRM-based filtering inspired by image manipulation detection methods [9], and the frequency branch analyzes the log magnitude spectrum obtained through FFT transformation to reveal spectral anomalies commonly associated with forged content [3]. An attention-based fusion module combines these complementary representations and adaptively emphasizes the most discriminative branch for each sample, following the motivation of prior multi-domain and multi-branch approaches [4], [5]. The model is trained and evaluated on the FaceForensics++ dataset using frame-level samples derived from video sequences [1]. Experimental results show that the proposed framework achieves a final test accuracy of 63.75%, demonstrating that multi-domain feature fusion is effective for improving deepfake detection performance. The results further indicate that attention-guided fusion helps the classifier exploit complementary forensic evidence beyond conventional RGB-only models.

DOI: http://doi.org/10.5281/zenodo.20199710

Ethnomedicinal Plants And Indigenous Knowledge System Of A Rural Community In Iligan City, Philippines

Authors: Samson L. Mangin, Edna B. Nabua

Abstract: Ethnomedicinal knowledge remains an important component of primary healthcare in many rural communities, yet it is increasingly threatened by modernization and environmental change. This study documents the ethnomedicinal plants used by the community of Sitio Langilanon, Barangay Pugaan, Iligan City, and examines their cultural and therapeutic significance. A mixed-methods research design was employed, integrating structured surveys and semi-structured interviews among selected settlers of Sitio Langilanon. Data collection focused on identifying medicinal plant species, ailments treated, methods of preparation, and associated indigenous health practices. Qualitative data were thematically analyzed, while quantitative data were summarized using descriptive statistics. Twenty-two (22) ethnomedicinal plant species were documented. Common methods of preparation included boiling, pounding, and infusion. These remedies were primarily used to treat ailments such as fever, diarrhea, cough, wounds, and hypertension. Findings revealed a strong community reliance on plant-based medicine and demonstrated the continued intergenerational transmission of indigenous medicinal knowledge. The study highlights the vital role of ethnomedicine in the local healthcare system of Sitio Langilanon. However, traditional medicinal practices face growing threats from cultural shifts and environmental degradation. Systematic documentation, cultural preservation initiatives, and scientific validation are essential to safeguard this knowledge and enhance its contribution to sustainable healthcare and broader scientific research.

DOI: http://doi.org/10.5281/zenodo.20200858

Biomertic Based MultiFactor Authentatication Using Behavioural Analysis

Authors: Ramineni Teja, Sankalp Bhawsar

Abstract: Passwords alone are no longer enough to keep systems secure in today’s rapidly changing cybersecurity land- scape. This study explores a smarter approach using biometric based multifactor authentication (MFA), focusing on how people interact with devices such as their typing patterns, mouse move- ments, and touch behavior. By combining these behavioral traits with machine learning, the system can continuously verify users with over 97 Percentage accuracy, without interrupting their experience. The research also looks at important challenges like preventing spoofing attacks, protecting user privacy, and ensuring the system can scale effectively. Overall, the findings suggest that behavioral biometrics can play a key role in building more secure and userfriendly authentication systems for the future.

DOI: http://doi.org/10.5281/zenodo.20201223

Smart Nutribot:A Web Application For Personalized Dietary Recommender Using Xgboost And Random Forest.

Authors: Jettty Varshitha, Gottipati Abhinaya, Kolusu Ankitha, Ms. S.A. Neelavani

Abstract: Smart Nutribot is an intelligent web- based application designed to provide personalized dietary recommendations using advanced machine learning techniques such as XGBoost and Random Forest. The system aims to address the growing need for customized nutrition plans by analyzing individual user data, including age, gender, weight, health conditions, dietary preferences, and lifestyle habits. By leveraging the predictive capabilities of ensemble learning models, the application generates accurate and tailored meal suggestions that promote healthy living and disease prevention. The web interface ensures user- friendly interaction, allowing users to input their details and receive instant recommendations in an accessible format. XGBoost enhances the model’s performance through efficient gradient boosting, while Random Forest improves robustness by reducing overfitting and increasing prediction accuracy. The integration of these models enables Smart Nutribot to deliver reliable and data-driven dietary guidance. Overall, this project demonstrates the practical application of machine learning in healthcare and nutrition, providing a scalable and efficient solution for personalized diet planning.

Comparative Study of Oxidative Stress Response Between Gram-positive and Gram-negative Bacteria

Authors: Sittie Fatmah M. Abdulrahman, Abdulraffy A. Alikhan, Najibah A. Casim, Janisah O. Hadji Amen, Samson L. Mangin, Junge B. Guillena, Doyne Grace Laparan, Sweet Maraesol Cabrera

Abstract: This study compared the oxidative stress responses of Gram-positive and Gram-negative bacteria exposed to hydrogen peroxide (5–20 mM) and a vitamin C–ferric chloride (FeCl₃) system under varying environmental conditions. Using the disc diffusion method, zones of inhibition were measured for Staphylococcus aureus, Streptococcus pyogenes, Pseudomonas aeruginosa, and Salmonella spp. Gram-negative bacteria were generally more susceptible to hydrogen peroxide, particularly under aerobic conditions, while Gram-positive bacteria showed greater resistance. The vitamin C–FeCl₃ system demonstrated consistent broad-spectrum antibacterial activity across all strains. Environmental factors, especially oxygen availability, influenced oxidative susceptibility. Findings highlight structural and physiological differences affecting bacterial oxidative stress tolerance and suggest the potential of vitamin C–FeCl₃ as an adjunct antimicrobial strategy.

DOI: https://doi.org/10.5281/zenodo.20201650

AI Based Crop Prediction And Recommendation

Authors: Sanskar Kadam, Varad Jamdar, Krushna Kapse, Snehal Shere, Shrawani Mule

Abstract: Farming is a big part of India's economy, making up about 17% of the country's GDP. It also helps support the lives of more than half of the people living in rural areas. But even with its importance, many Indian farmers face crop failures every year. This often happens because they don't have access to affordable tools that can give them good advice on which crops to plant, based on scientific research. To solve this problem, we've developed a system that uses artificial intelligence to predict and recommend crops. It combines special hardware that senses the condition of the soil with a type of machine learning model that considers many factors at once. This allows the system to give farmers personalized advice in real time, helping them make informed decisions about which crops to plant. Here's a rewritten version of the input text in a more human-like tone, similar to the provided reference human samples: When it comes to measuring soil nutrients, we've developed a hardware prototype that's pretty impressive. It's made up of an RS-485 Modbus NPK sensor, a DHT11 temperature and humidity sensor, a capacitive soil moisture sensor, and an Arduino UNO microcontroller. In the early stages, we even experimented with a TDS sensor as a low-cost alternative for estimating soil nutrient characteristics. This approach helped us keep costs down while still testing the system's architecture. But in the end, we decided to use actual NPK sensor readings as the primary input for soil nutrients. We also trained and compared seven different machine learning models using a dataset of 2,200 agricultural samples covering 22 different crop classes. These models included Random Forest, XGBoost, LightGBM, SVM, Gradient Boosting, KNN, and Logistic Regression. And what we found was that a soft-voting ensemble combining Random Forest, XGBoost, and LightGBM achieved an impressive 99.77% test accuracy and 99.73% mean cross-validation accuracy. But here's the thing: we didn't just stop at soil nutrients. We also incorporated real-time weather data for temperature, humidity, and rainfall into our model, using the OpenWeatherMap API. This allows us to provide location-aware recommendations that take into account the specific weather conditions in a given area. And the best part? The entire system is deployed as a user-friendly Gradio web application, with three different output tabs: crop recommendation with confidence bars, soil health analysis with fertiliser advice, and a seasonal crop planner. What's really exciting about this system is that it directly supports the United Nations' Sustainable Development Goal 2 – Zero Hunger. By providing farmers with accurate and reliable recommendations, we can help increase crop yields and reduce hunger around the world. It's a big goal, but we're hopeful that our system can make a real difference.

DOI: https://doi.org/10.5281/zenodo.20204819

Self-Attention–Driven Vision Transformer Model For Autism Identification And Cognitive Skill Enhancement

Authors: M. Menakapriya, M. Rasika, G. Prabha, R. Lavanya, S. Mounika, V. Nagarajan

Abstract: Early identification and treatment of Autism Spectrum Disorder (ASD) pose difficulties due to its complex neurological nature and heterogeneous symptoms. In this work, an innovative Self-Attention-Driven Vision Transformer (SA-ViT) model is introduced to cater to both ASD detection and cognitive skills improvement within one approach. Our work benefits from self-attention properties of vision transformers to detect subtle patterns in the behavior of ASD individuals as well as generate structured cognitive stimulation material. By extracting features from facial imagery, videos, and behavioral data, our SA-ViT can classify ASD samples with 97.6% accuracy on the ASD Facial Image Dataset, which beats regular CNN models (91.2%) and regular ViTs (94.8%). In terms of cognitive skills improvement, we were able to develop personalized structured tasks that resulted in 34.2% improved visual memory retention and 28.7% enhanced pattern recognition after eight weeks. The use of explainable AI approaches (Grad-CAM) enhances our system's applicability in a medical setting.

DOI: https://doi.org/10.5281/zenodo.20205889

Real-World Case Study: Evaluating AI-Powered And Traditional Signature Approaches To Email Phishing Threats

Authors: Shrushti Kaza, Akhila Harshini Gadamsetty, Abhijeet Raj, Pranav Veer Singh, Dr M Umamaheswari

Abstract: Email-based phishing is among the persistent and costly cybersecurity challenges which exploit human gullibility, social engineering, and organizational frailties for gaining un-approved access to sensitive information. Signature-based cyber-security strategies use preset patterns, blacklists, and heuristic approaches to identify phishing emails. While signature-based detection systems can recognize phishing emails with known char-acteristics successfully, they usually fail to identify sophisticated attacks which evade recognition due to their novelty or disguise. On the other hand, modern technologies based on ML and NLP employ numerous features including email body, natural language used, sender behavior, URLs embedded in an email, and additional metadata. The ability of such approaches to generalize makes them applicable in the detection of previously unseen phishing campaigns. In this study, comprehensive comparison between AI-powered phishing detectors and traditional signature-based methods is conducted using a hand-curated dataset with both legitimate and malicious samples of emails. Criteria for evaluation include detection rates, false positives and negatives, as well as computational resources consumed. The experiments show that AI-based techniques outperform traditional systems in terms of recognizing unknown phishing emails. However, superior performance comes at the expense of greater computational loads and increased requirements for tuning and maintaining AI models. Also, this study provides practical guidance for integrating AI-based phishing detectors into corporate email systems, considering deployment issues, scaling, and computational resources needed. Based on the experiment results presented in the paper, recommendations are made regarding implementation of AI solutions for phishing attacks.

DOI: http://doi.org/10.5281/zenodo.20205928

Ai Prompt Helper

Authors: M.Benita Roy, L.Bhuvaneswar, K.V.Bharath Kumar, K.Vishnuvardhan

Abstract: The proliferation of digital content creation has created significant challenges for content creators, educators, and professionals who require efficient prompt engineering and AI-assisted content generation workflows. This paper proposes AI Prompt Helper, an intelligent system designed to optimize prompt construction, enhance content generation quality, and automate repetitive workflows for various AI applications. The proposed system integrates advanced natural language processing techniques, user-friendly interface design, and intelligent automation mechanisms to provide users with real-time suggestions, template management, and workflow optimization. The paper presents the system architecture, implementation details, performance evaluation, and practical use cases demonstrating the effectiveness of AI Prompt Helper in improving content generation productivity by up to 65% and reducing user cognitive load. Further, this paper provides insights into the system's design philosophy, technical implementation, and future research directions for intelligent prompt optimization systems.

AI-Powered Resume Analyzer & Smart Job Matching Platform

Authors: Vibhuti Chaddha, Nikunj Agarwal, Dr. Yatu Rani

Abstract: Abstract: The rapid growth of digital recruitment platforms and online hiring systems has significantly increased the volume of job applications received by organizations, making manual resume screening and candidate shortlisting time-consuming and inefficient. Traditional recruitment methods and keyword-based Applicant Tracking Systems (ATS) often fail to accurately identify suitable candidates due to limited semantic understanding and dependency on exact keyword matching. To address these limitations, this paper presents SmartHireX: AI-Powered Resume Analyzer & Smart Job Matching Platform, an intelligent recruitment system designed to automate resume analysis, improve candidate-job matching accuracy, and enhance recruitment efficiency. The platform is designed to support both recruiters and job seekers. Recruiters can automate candidate screening, ranking, and shortlisting processes, thereby reducing manual effort and improving hiring decision-making. Job seekers benefit from personalized resume improvement suggestions, missing keyword identification, skill gap analysis, and intelligent job recommendations that help optimize resumes according to industry and ATS standards.The system is implemented using modern web and AI technologies, including React and Tailwind CSS for frontend development, Flask (Python) for backend services, spaCy for Natural Language Processing tasks, and scikit-learn for Machine Learning and similarity analysis. Supabase is used for database management, while deployment can be supported using scalable cloud infrastructure. Experimental analysis and system evaluation demonstrate that the proposed platform improves recruitment accuracy, reduces screening time, and enhances candidate-job compatibility analysis compared to traditional recruitment approaches.

DOI: https://doi.org/10.5281/zenodo.20215014