Authors - Mohammad Nasar, Mohammad Abu Kausar, Md. Abu Nayyer Abstract - Smart farming technologies, particularly image-based crop disease detection systems, are significantly transforming agriculture by enabling farmers to respond promptly to emerging threats. However, deploying centralized AI models in remote or rural farming communities remains challenging due to limited internet connectivity, heterogeneous data sources, and privacy concerns. This paper presents a federated learning (FL) framework designed to support secure and privacy-preserving crop disease detection in decentralized agricultural environments. In the proposed approach, lightweight convolutional neural networks (CNNs) are trained locally on crop images captured by drones or field sensors, ensuring that sensitive data remains at the source. Instead of transmit-ting raw data, only model updates are shared with a central server, which aggregates them using the Federated Averaging (FedAvg) algorithm. To enhance security and model integrity, a cosine similarity-based filter is incorporated to detect and discard malicious or corrupted updates. Additionally, an attention-based heatmap module is introduced to visually highlight disease-affected regions in the crop images, improving system transparency and building user trust. Experimental evaluations on the PlantVillage dataset and real-world drone imagery of wheat and tomato crops demonstrate that the framework achieves an average 4.8% improvement in F1-score under non-IID data conditions com-pared to centralized models, alongside a 9.3% reduction in communication overhead at 50 nodes. This solution is well-suited for resource-constrained rural settings and contributes to scalable, intelligent, and sustainable agricultural digitization.
Authors - Makrand Kadam, Prashant Dhotre Abstract - Active Directory (AD), created by Microsoft, plays a central role in how most organizations manage user identities, access, and system permissions. Because of its importance in controlling access across enterprise networks, it’s a prime target for cyber attackers. This paper takes a closer look at how AD is structured, highlighting key vulnerabilities and common methods used by attackers—like stealing credentials, escalating privileges, using Pass-the-Hash, or creating Golden Tickets. It also examines real-life security breaches to show just how critical AD security has become. Beyond identifying threats, the paper outlines practical defenses such as improved auditing, network segmentation, hardening of domain controllers, and adopting zero trust models. By bringing together current research and best practices from the field, this study offers cybersecurity professionals and system administrators a clear, actionable guide to better protect their AD setups against today’s increasingly sophisticated threats.
Authors - Rajeshwari S B, Sajini G, Jagadish S Kallimani Abstract - In today's fast-paced world, people frequently experience high levels of emotional stress and struggle to access personalized mental health support. Traditional mental health services often face challenges such as limited accessibility and a generalized approach, which can overlook individual emotional nuances. This one-size-fits-all method may fail to address the specific needs of individuals. The need for more responsive and individualized mental health resources has never been greater, highlighting the importance of innovative solutions that offer contextually relevant and personalized care. The Srimad Bhagavad Gita, a timeless spiritual text, offers profound teachings on managing emotions, finding purpose and maintaining balance in life. This project aims to develop a Mental Health Conversational Assistant that provides users with motivational responses rooted in the wisdom of the Bhagavad Gita. This chatbot lever-ages NLP and NLU to detect both the emotional state and the intent behind user queries. It uses advanced techniques like emotion detection and semantic understanding to analyze user input and respond with relevant verses from the Bhagavad Gita. By integrating Hugging Face's emotion detection models and sentence-transformer embeddings, chatbot is capable of providing tailored, contextually relevant verses from Bhagavad Gita that match both the user’s emotions and the meaning of their queries.
Authors - Reena S. Satpute, Nura Muhammad Sani Abstract - Effective disease surveillance and outbreak response are high public health priorities, particularly in resource-limited healthcare systems facing significant challenges with data collection, predictive modeling, and coordinated intervention strategies. It hence proposes an integrated AI-driven framework responding to these challenges and strengthening the capacity of public health authorities to better manage infectious disease outbreaks. This framework will integrate real-time data from different sources such as electronic health records, syndromic surveillance, environmental indicators, and social media in a harmonious way into one centralized repository for the enablement of advanced predictive modeling techniques. It mainly comprises a disease forecasting model, an outbreak detection algorithm, and an intervention optimization tool integrated to provide early warning, accurate prediction, and data-driven decision support. An extensive performance evaluation of the framework was conducted using accuracy, sensitivity, specificity, and area under the ROC curve as some of the metrics. The results are very satisfactory, while the overall accuracy is 94.8% (95% CI: 93.2%-96.4%) with an AUC-ROC of 0.967 (95% CI: 0.951 - 0.983). Comparative performance analysis using state-of-the-art deep learning architectures further establishes the superiority of the proposed hybrid model. Transparency and explainability in the framework were guaranteed up to an extent that allows understanding the models and decision-making process by the public health authorities.
Authors - Adulfo R. Arevalo, Ronald L. Pancho Abstract - This study examines the forecasting performance of three prominent models— Autoregressive Integrated Moving Average (ARIMA), Support Vector Regression (SVR), and Random Forest (RF)—in predicting the short-term price movements of Bitcoin (BTC) and XRP using real daily closing price data from January to June 2024. The research aims to determine which model delivers the most accurate forecasts by applying each technique to actual market data and evaluating their predictive outputs using three standard accuracy metrics: Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Squared Error (RMSE). Results show that ARIMA consistently outperformed both SVR and Random Forest in forecasting accuracy, recording the lowest average errors across all metrics for BTC and XRP. While SVR performed poorly in capturing XRP's volatility, Random Forest showed moderate results but lagged behind ARIMA. These findings highlight ARIMA's robustness and practicality as a forecasting tool for short-term cryptocurrency price movements. The study concludes with model-specific recommendations, encouraging the use of ARIMA for near-term price prediction and suggesting further exploration into hybrid and deep learning models for future research.
Authors - Rajeshwari S B, Jagadish S Kallimani Abstract - In the realm of human-computer interaction, the development of input devices accommodating diverse user needs is paramount. Traditional peripherals like mice and keyboards, though effective in typical settings, often falter in specialized conditions such as low-light environments. Addressing these constraints, this study introduces a novel gesture and voice-controlled mouse system. By leveraging real-time video and audio inputs, the system interprets user gestures and voice commands, eliminating reliance on conventional input tools. Additionally, it tackles unresolved interface design issues, striving to optimize user experience and broaden computing resource accessibility. Key to its functionality is the integration of MediaPipe for precise gesture recognition and natural language processing for voice command interpretation. Utilizing standard webcam and microphone hardware enables hands-free cursor control and command execution, particularly aiding users with mobility limitations. Methodologically, the system blends foundational theoretical principles with innovative design concepts for practical realization. Rigorous testing demonstrates a 90% accuracy rate in gesture recognition and a notable 60% accuracy in low-light conditions. Voice command interpretation achieves an impressive 95% accuracy across various accents and speech patterns, with a significant reduction in task completion time com-pared to conventional setups. This research underscores the potential of an affordable and adaptable technology, promising enhanced digital accessibility through ubiquitous webcam and microphone integration.
Authors - Thanh D. Nguyen, Anh T. N. Nguyen Abstract - This study examines how habitual behavior and users’ perceptions of security influence their intention to adopt and use e-wallet services. Building upon the original Technology Acceptance Model (TAM), the research integrates elements from both the Unified Theory of Acceptance and Use of Technology (UTAUT) and its extended version, UTAUT2. Data were collected from 265 valid respondents via an online survey and analyzed using reliability testing, exploratory factor analysis (EFA), confirmatory factor analysis (CFA), and structural equation modeling (SEM). The findings reveal that perceived security, general privacy concerns, and habit all positively affect the intention to adopt e-wallets. Further-more, this intention significantly influences actual usage behavior.
Authors - Nureni Ayofe Azeez, Solomon Akinboro, Isiekwene Chinyere Chioma, Charles Van der Vyver, Awoseemo Babajide Abstract - The propagation of fake news in the digital age has raised momentous concerns about the integrity of information dissemination. This research aims to carry out a comparative assessment of machine learning algorithms for identifying fake news in social media. To achieve this, an attempt was made to collect a wide-ranging dataset of news articles, encompassing both real and fake sources. In the initial phase of the research, we preprocess the news data, extracting relevant features such as text content, metadata, and source credibility. Leveraging a wealth of labelled data, these models learn to differentiate between genuine and fake news articles, while also providing insights into the characteristics that contribute to their classification. In a world where information authenticity is paramount, the implementation of this system provides a valuable tool for discerning consumers of news and enables the swift identification of misinformation sources. The models were trained using four different datasets. Several machine learning algorithms were used for carrying out comparative assessment in terms of their efficiency and accuracy to detect fake news. Selected metrics were chosen for a comparative assessment. At the end of the experiments, for Dataset 1, the Passive Aggressive Algorithm had the highest accuracy of 0.8333, while for Dataset 2, the Passive Aggressive Algorithm had the highest accuracy of 0.914285. However, for Dataset 3, the AdaBoostClassifier Algorithm had the highest accuracy of 0.9142857. Furthermore, in Dataset 4, several algorithms which including Logistic Regression, Passive Aggressive, Ridge Classifier, SGD Classifier, AdaBoost Classifier, Bagging Classifier, Extra Tree Classifier, Gradient Classifier, Random Forest, NUSVC, Linear SVC, Decision Tree, BernoulliNB, and GaussianNB, all had the highest accuracy of 1.0000000. A bar chart was used to represent the results graphically, showing the minimum, maximum, and quartile range.
Authors - Akash Karale, Pratvina Talele Abstract - With the rapid advancement of deepfake audio technology, producing artificial voices that sound remarkably realistic is now simpler than ever. As this becomes available In addition to opening up new creative possibilities, it also poses grave security, disinformation, and fraud concerns. Although many of the deepfake audio detection tools available today are accurate, their application in real-time situations is limited by their high processing overhead. In this work, we introduce a Convolutional Neural Network (CNN)-based lightweight deepfake audio detection model. Our model can detect AI-generated speech accurately and with minimal processing demands by utilizing Mel-Spectrogram features and an effective CNN architecture. The model works well with high accuracy and low latency, according to tests on standard datasets, which makes it appropriate for real-world uses like media verification and voice authentication. Looking ahead, we aim to improve the model's ability to withstand adversarial attacks and adapt to new types of deepfake generation techniques. In the future, we would like to improve the model's robustness against hostile attacks and make it more flexible to accommodate new deepfake generation techniques
Authors - Ram Joshi, Rutuja Khedkar, Prema Sahane, Shantanu Totewad, Nikhil Tarate, Prajodeep Patil, Kedarnath Zunjare Abstract - Deepfakes are manipulated or AI-generated images and videos that appear authentic but are not. A decade ago, detecting deepfakes was relatively straightforward due to the limitations of the technology. However, recent advancements in AI, machine learning, and deep learning have made deepfakes nearly undetectable. Although this technology is often used for entertainment, it is increasingly being used for malicious purposes. The widespread availability of AI tools has further facilitated the creation of convincing deepfakes by fraudsters. Given that people tend to trust what they see and hear, they are particularly susceptible to the influence of realistic deepfakes.
