Authors: Assistant Professor Gargi Mishra, Jayarekha.B

Abstract: Dietary deficiencies continue to be one of the most prevalent modifiable risk factors for metabolic disorders despite the challenges associated with traditional self-reporting techniques. This paper outlines an architecture for smart nutrition monitoring using a combination of wearable multimodal sensors and mHealth solutions. In particular, the architecture relies on signals collected from a wrist-wearable device, acoustic signals from an earable sensor, and annotated meal information from a mobile app to recognize eating events and identify the composition of each meal. A lightweight edge classifier processes the signal streams, whereas the mobile app provides feedback and a summary of nutritional data over time. The methodology covers the collection, preprocessing, detection of eating events, categorization of food categories, and visualization components. An analysis of operational performance of the proposed framework shows that a multimodal approach is better at lowering false positives in eating event recognition than a unimodal approach, and also allows users to stay motivated by providing feedback through a mobile application.

DOI: https://doi.org/10.5281/zenodo.24136527

Published by: vikaspatanker

Authors: Assistant Professor Dr R. Karthikeyan, Assistant Professor Ms. M. Sivapriya

Abstract: The increasing speed and complexity of modern cyber threats have brought into light the limitations of the traditional and human-centred processes in dealing with these threats. In this paper, we propose an innovative Cyber Defence Framework based on Agentic Artificial Intelligence (AACDF), which involves multi-agent orchestration along with autonomous decision-making abilities in the whole process of detecting and responding to threats. In particular, this framework uses dedicated agents such as perception, reasoning, action, and coordination agents working in a hierarchical manner inside the control loop based on the MAPE-K model enriched with large language model reasoning. We discuss the architecture design, the interaction mechanism between agents, and the decision-making process regarding autonomous actions for containing and remediating detected threats. A comparative analysis of the proposed framework with state-of-the-art autonomous cyber defence approaches highlights its advantages in terms of explainability and adaptability.

DOI: https://doi.org/10.5281/zenodo.24136425

Published by: vikaspatanker

Authors: Olotuah Roseline Temilade, Professor Oluwadare Samuel

Abstract: Accurate iris edge detection and boundary localization is seen as the prerequisite for high-performance biometric authentication systems. Traditional edge detection operators like, Canny, Sobel, Prewitt and feature-engineered shallow machine learning models are often beset with limitations such as severe sensitivity to noise, specular reflections, eyelid occlusions, and acute class imbalance when predicting thin boundary pixels. This paper presents a novel deep learning-driven ensemble framework that reformulates iris edge detection as a region-based binary semantic segmentation task using 2,639 high-resolution grayscale images (320 × 280 pixels, from the Standardized CASIA-Iris Interval benchmark database. The study adopts a Design Science Research (DSR) methodology, structured into three sequential phases: automated boundary localization for pseudo ground-truth generation using Circular Hough Transform, deep convolutional neural network segmentation (U-Net, ResNet-UNet, and Attention U-Net) and advanced ensemble mechanisms which includes Soft Voting, Stacking Meta-Learning, and Hybrid Fusion incorporating classical gradient maps. Raw grayscale iris images were made to undergo spatial standardization, Gaussian smoothing, and intensity normalization. The preprocessed tensors are passed concurrently through standalone CNN models and classical edge pipelines. The probability maps generated by CNNs are aggregated via soft voting, stacking, and hybrid fusion, and subsequently fused with classical gradient maps to derive precise binary segmentation masks and sharp boundary edges for improved robustness. The comparative evaluation, was done using four performance metrics; Intersection over Union (IoU), F1-Score, Precision and Recall. Experimental results demonstrate that deep learning architectures significantly outperform classical edge operators. Furthermore, ensemble models achieve superior segmentation dependability and boundary precision, attaining an Intersection over Union (IoU) of 0.9412, F1-Score of 0.9697, Precision of 0.9723, and Recall of 0.9671. The proposed region-based ensemble paradigm effectively allays pixel class imbalance, demonstrates superior robustness against intra-class variations, and establishes an accurate benchmark, especially, for next-generation biometric identification pipelines.

DOI: https://doi.org/10.5281/zenodo.24134661

Published by: vikaspatanker

Authors: Dhobale Kartik Rajendra, Prof. N. G. Rokade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Driver fatigue is one of the most common and preventable causes of road accidents. Long hours, monotonous highways and lack of rest reduce alertness, often without the driver noticing. Many existing solutions depend on wearable physiological sensors or expensive vehicle hardware that most drivers do not have. This paper reviews driver drowsiness detection approaches and presents the design of a real-time system that needs only an ordinary webcam. Behavioural, physiological and vehicle-based methods are compared, and vision-based techniques are examined in detail, including face detection, dense facial landmark extraction with MediaPipe Face Mesh, eye aspect ratio and mouth aspect ratio measures, PERCLOS and convolutional neural networks that verify eye closure and yawning. Public datasets for drowsiness and yawning are discussed. In the proposed design, OpenCV captures video, Face Mesh locates 468 landmarks, ratios are computed in every frame, and a CNN confirms closed eyes and yawns to reduce false alarms; if signs persist across consecutive frames, an audible alarm is raised. Research gaps, objectives, a comparison of methods and challenges such as night-time lighting, spectacles, head pose and individual variation are discussed.

DOI: https://doi.org/10.5281/zenodo.24127999

Published by: vikaspatanker

Authors: Kokane Sanika Santosh, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Deepfakes are realistic manipulated images, videos and audio created with deep learning and generative models that can imitate a person's face, voice, expressions or movements. Although useful in entertainment and education, they are misused for misinformation, fake identities, fraud and reputational harm. This paper reviews AI-based deepfake detection across both visual and audio modalities. It explains generation methods, then examines visual detection from facial texture, lighting, eye and lip movement and inter-frame differences using convolutional and attention-based networks; generalizable detectors for fully synthetic images; audio detection from spectral and prosodic voice characteristics using anti-spoofing models; and audio–visual methods that detect mismatches between speech and lip motion. Benchmarks including FaceForensics++, ASVspoof and FakeAVCeleb are discussed. The detection process of collecting and preprocessing media, extracting features and applying machine learning or deep learning models is formalized in a five-layer architecture. Research gaps, objectives, a comparison of modalities and challenges such as generalization, compression, adversarial evasion and explainability for forensic use are presented. The review concludes that combining audio and visual evidence gives the most reliable detection for social media, journalism, banking, law enforcement and digital forensics.

DOI: https://doi.org/10.5281/zenodo.24127827

Published by: vikaspatanker

Authors: Mule Isha Bhagvant, Prof. R. V. Varhade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Intelligent Tutoring Systems (ITS) aim to give every student the benefits of one-to-one tutoring by using Artificial Intelligence to understand individual needs, abilities and progress. This paper reviews AI-powered ITS. It explains the classic architecture of domain, learner, pedagogical and interface models, and surveys learner modelling through knowledge tracing with Bayesian and deep recurrent approaches, adaptive selection of content and practice, automated feedback and hints, and conversational tutors from early dialogue systems to tutors built on large language models. Evidence on effectiveness is examined, including meta-analyses comparing ITS with human tutoring and classroom instruction, and recent studies of language-model-generated hints. A five-layer AI tutoring architecture is proposed, comprising a content and domain layer, a learner model, an adaptive engine, an AI chatbot tutor and a teacher and parent dashboard with progress reports, quizzes and practice tests. Research gaps, objectives, a comparison of tutoring approaches and challenges including accuracy of generated explanations, over-reliance, privacy, equity of access and multilingual support are discussed. The review concludes that AI tutors can deliver personalized, interactive and effective learning when designed with sound pedagogy and teacher involvement.

DOI: https://doi.org/10.5281/zenodo.24127437

Published by: vikaspatanker

Authors: Nalawade Tanvi Dnyaneshwar, Prof. J. M. Patade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Traffic congestion is a major urban problem, worsened by conventional traffic lights that operate on fixed time intervals without considering how many vehicles are waiting on each approach. This paper reviews approaches to adaptive traffic signal control and focuses on using machine learning to predict suitable green-light durations for each lane of an intersection based on traffic demand. It traces the field from classical signal timing formulas through adaptive systems that adjust timings from detector data, to data-driven prediction of traffic flow and reinforcement learning controllers that learn signal policies through simulation. Input parameters such as lane-wise vehicle counts from cameras, traffic density, queue length, waiting time and time of day are discussed, along with traffic simulators used for safe training and evaluation. A five-layer prototype architecture is proposed in which vehicle detection feeds feature construction, a regression model predicts green times, a safety and constraint layer enforces minimum and maximum phases and pedestrian crossing times, and a controller updates the signal. Research gaps, objectives, a comparison of control approaches and challenges such as sensor reliability, mixed Indian traffic and coordination across intersections are discussed.

DOI: https://doi.org/10.5281/zenodo.24127044

Published by: vikaspatanker

Authors: Dr. Madala Guru Brahmam, M. Priyanka, D. Nakshatra Maanasa, M. Likhitha, K. Annapoorna

Abstract: Career planning can be challenging when a learner has a goal but is unsure which skills to develop, which resources to use, or how to organise the work over time. Relevant infor-mation is often spread across learning platforms, documentation sites, progress trackers, and job portals. This paper presents FutureEra, a web-based career-planning platform designed to bring these activities into one workflow. Users can enter a career goal and profile details, receive a structured roadmap generated with a large language model, explore milestone-related learning resources, record daily activity, compare career paths, use peer-benchmarking and community features, and interact with an AI mentor. The platform is implemented using React and Vite on the frontend, Node.js and Express.js on the backend, and MongoDB for persistent storage. Its generative features use the Groq API with an OpenAI GPT-OSS model selected through the server environment. The application also provides account management, password recovery, progress summaries, and PDF export. Generated roadmaps and estimates are presented as plan-ning suggestions, not as guarantees of salary, completion time, or employment. This paper describes the motivation, related work, system architecture, implementation, current limitations, and an evaluation plan. No quantitative performance or user-study results are claimed because verified experimental measurements are not available for this study.

DOI: https://doi.org/10.5281/zenodo.24126985

Published by: vikaspatanker

Authors: Padir Aditi Vinod, Prof. J. M. Patade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Machine learning models increasingly influence decisions in healthcare, finance, hiring and public services, yet many of the most accurate models are complex black boxes whose reasoning cannot be inspected. Explainable Artificial Intelligence (XAI) aims to make these decisions transparent and interpretable so that users can understand, trust, debug and contest them. This paper reviews XAI as a general discipline rather than for a single domain. It clarifies concepts of interpretability, explainability, transparency and faithfulness, and organizes methods into inherently interpretable models, global explanations such as partial dependence and accumulated local effects, local feature attribution with LIME, SHAP and integrated gradients, visual explanations for deep networks, concept-based explanations and counterfactuals. Evaluation of explanations, including sanity checks that reveal misleading saliency maps, and insights from social science on what makes a good explanation are discussed. A five-layer explainable ML pipeline comprising data, model, explanation engine, evaluation and audience-specific presentation is proposed, with research gaps, objectives, a comparison of methods and challenges such as the accuracy–interpretability trade-off, unstable explanations and regulation. The review concludes that explanations should be faithful, evaluated and designed for the people who use them.

DOI: https://doi.org/10.5281/zenodo.24126787

Published by: vikaspatanker

Authors: Yashasvi Sharma, Amrish Kumar

Abstract: Signature-based security tools detect attacks they already know and are blind to zero-day attacks, while many machine learning intrusion detection systems still suffer from high false-positive rates and have been evaluated mainly on static, pre-recorded datasets. This paper presents an intelligent threat detection system that learns normal network behaviour, flags deviations from it, and combines supervised classifiers with an anomaly detector. A synthetic network-flow generator produces normal traffic, three known attack types (DoS flood, port scan, SSH brute force) and three attack types withheld from training (Slowloris, botnet beacon, data exfiltration). Four detectors are compared over five random runs on 18,000 test flows: signature rules, a Random Forest, an MLP neural network, and a hybrid of Random Forest, MLP and an Isolation Forest fitted on normal traffic. The hybrid detects 100% of known attacks and 68.4% of unseen attacks, against about 90% and 0.1% for signature rules, at a 3.0% false-alarm rate. Its gain is uneven: exfiltration is detected only 5% of the time. The results come from author-generated synthetic data, so they show the behaviour of the design rather than field performance. CICIDS2017, a CNN-LSTM model, SHAP explanations, detection-latency measurement and live traffic remain to be done and are not claimed. The contribution is a design and evaluation framework for detecting known and unknown attacks, not a new algorithm.

DOI: https://doi.org/10.5281/zenodo.24126719

Published by: vikaspatanker