Authors: Kurhade Suyash Rajaram, Prof. R. V. Varhade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Phishing emails deceive individuals into revealing credentials and personal data, and attackers can now use large language models to write fluent, personalized phishing messages at scale that slip past rule-based filters. This paper reviews AI techniques for an adaptive phishing email detector that analyses message text, sender information and embedded URLs in real time. It traces detection from structural and feature-based approaches through natural-language analysis of intent and urgency to transformer models such as BERT and RoBERTa fine-tuned for phishing classification. URL analysis with deep character models, sender-authentication checks and behavioural signals are reviewed. Particular attention is given to the new threat of AI-generated phishing, including studies showing that language models can craft convincing spear-phishing emails and that the same models can help detect them. A five-layer architecture comprising email ingestion, multi-view feature extraction, transformer-based and URL classifiers, risk fusion with explanation, and user interaction with feedback is proposed. Research gaps, objectives, a comparison of models and challenges including adversarial adaptation, false positives, privacy and multilingual phishing are discussed. The review concludes that adaptive, explainable, multi-view detection is necessary as phishing becomes AI-assisted.

DOI: https://doi.org/10.5281/zenodo.24035093

Published by: vikaspatanker

Authors: Khot Shruti Rahul, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Clinical diagnosis rarely relies on one source of information: doctors combine medical images, symptoms described by patients, clinical reports, laboratory results, electronic health records, and sometimes audio or video such as heart sounds or gait recordings. Multimodal Artificial Intelligence integrates such heterogeneous data to improve the accuracy, efficiency and reliability of diagnosis. This paper reviews multimodal AI for intelligent healthcare diagnosis assistance. It introduces the taxonomy of multimodal machine learning, covering representation, alignment and fusion, and compares early, intermediate and late fusion strategies. Applications are reviewed in disease detection, image interpretation with reports, personalized treatment and patient monitoring, together with vision–language models that learn from paired images and text and generalist medical foundation models that accept many input types. Public multimodal datasets are discussed. A five-layer architecture comprising multimodal data acquisition, modality-specific encoders, fusion, diagnostic reasoning and clinician-facing explanation is proposed, with research gaps, objectives, a comparison of fusion strategies and challenges including missing modalities, data alignment, privacy, interpretability and validation. The review concludes that multimodal AI can provide more comprehensive patient analysis and faster decisions when it is rigorously validated and clinically supervised.

DOI: https://doi.org/10.5281/zenodo.24034883

Published by: vikaspatanker

Authors: Bhor Darshan Vasant, Prof. M. A. Auti, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Edge Artificial Intelligence (Edge AI) combines machine learning with edge computing so that data are processed on or near the devices that produce them, such as sensors, cameras, smartphones, IoT nodes and embedded systems, rather than entirely in remote cloud servers. This brings decisions closer to the data source, reducing latency, bandwidth use and dependence on connectivity while improving privacy. This paper reviews the foundations and practice of Edge AI for real-time applications. It examines the edge computing paradigm and edge intelligence levels, the partitioning of neural networks between device, edge server and cloud, efficient architectures such as depthwise-separable networks, early-exit inference, and compression by pruning, quantization and knowledge distillation. Tiny machine learning on microcontrollers and standard benchmarks for measuring edge inference are discussed. Applications in smart surveillance, industrial monitoring, healthcare wearables, autonomous robots and smart cities are summarized. A layered device–edge–cloud architecture is proposed, together with research gaps, objectives, a comparison of deployment tiers and an analysis of challenges including limited resources, heterogeneity, security and model updates. The review concludes that Edge AI is essential for low-latency, private and reliable intelligent systems.

DOI: https://doi.org/10.5281/zenodo.24034625

Published by: vikaspatanker

Authors: Tirandaj Fatima Khalid, Prof. S. S. Bhor, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Artificial intelligence, and generative modelling in particular, is creating new opportunities in bioinformatics. Conventional computational biology relied on physics-based simulation, sequence alignment and manual feature engineering, each limited by computational cost and the need for expert interpretation. This paper reviews how deep learning and generative models are used in bioinformatics research. It first examines protein structure prediction, where deep learning achieved a step change in accuracy and extended to complexes of biomolecules, and discusses what such predictions can and cannot tell researchers. It then reviews representation learning on biological sequences, in which transformer models trained on large public sequence databases learn representations that capture structural and evolutionary signal and support downstream tasks such as property and variant-effect prediction. The paper surveys the families of generative models used in the field, including variational autoencoders, graph neural networks, diffusion models and transformers, at the level of their modelling assumptions, and reviews published applications in computational drug discovery such as virtual screening and candidate prioritization. It discusses data resources, evaluation practice, the central role of experimental validation, and the governance of research, including responsible-use and biosecurity frameworks. Research gaps, objectives, a comparison of model families and challenges are presented. The review concludes that these methods are decision-support tools for researchers and do not replace laboratory experimentation or institutional oversight.

DOI: https://doi.org/10.5281/zenodo.24033350

Published by: vikaspatanker

Authors: Pabale Sakshi Fakir, Prof. J. M. Patade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Earthquakes, floods, landslides and building collapses cause serious damage and loss of life, and locating trapped or stranded people quickly is one of the most critical tasks for rescue teams. Ground searches are slow and often dangerous in unstable or inaccessible areas. AI-powered drones offer a faster and safer alternative by combining aerial mobility with onboard cameras, thermal sensors, GPS and computer vision. This paper reviews the design of AI drones for disaster rescue. It examines aerial datasets and deep object detectors for recognizing people, vehicles and damaged structures from above, thermal imaging for finding victims at night or under light cover, coverage path planning for searching large areas systematically, simultaneous localization and mapping for operation where GPS is unreliable, and drone networks that relay communication when ground infrastructure fails. Practical constraints of battery life, payload, onboard computing and regulation are discussed. A five-layer architecture comprising sensing, onboard edge AI, navigation and control, communication and a ground control station is proposed, followed by a research gap analysis, objectives, comparison of detection models and a discussion of challenges. The review concludes that AI drones can cover large areas quickly, reach inaccessible locations and reduce risk to rescuers, making disaster response faster, safer and more effective.

DOI: https://doi.org/10.5281/zenodo.24033116

Published by: vikaspatanker

Authors: Bhalerao Onkar Rohidas, Prof. M. A. Auti, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Autonomous vehicles must predict how nearby vehicles, cyclists and pedestrians will move within the next few seconds in order to plan safe actions. Sending raw sensor data to the cloud for this prediction introduces latency and dependence on network connectivity that can jeopardize safety. This paper reviews Edge AI approaches that perform perception and motion prediction on onboard embedded hardware. It surveys trajectory forecasting models, from social pooling of recurrent networks and generative adversarial prediction to graph-based models that encode high-definition maps and agent interactions, and multi-modal predictors that output several likely futures. Large-scale motion forecasting datasets are discussed. The review then examines techniques for running these models within strict latency and power budgets, including pruning, quantization to integer arithmetic, knowledge distillation and hardware–software co-design on automotive accelerators. A five-layer onboard architecture covering sensing, perception, edge prediction, planning and vehicle control is proposed, with research gaps, objectives and a comparison of prediction models. The review concludes that compact, quantized multi-modal predictors running at the edge can deliver the low-latency, reliable forecasts that safe autonomous driving requires.

DOI: https://doi.org/10.5281/zenodo.24032970

Published by: vikaspatanker

Authors: Gawade Aditi Dattatray, Prof. C. A. Kolse, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Digital platforms, cloud services, online transactions and data-driven applications collect and process large volumes of personal and confidential information every day, increasing the risk of data breaches, unauthorized access, identity theft and misuse. Protecting privacy while still allowing data to be used for legitimate purposes has therefore become a core requirement of modern systems. This paper reviews the principal privacy-preserving technologies and how they can be combined for secure data management. It covers data encryption at rest and in transit, anonymization models such as k-anonymity and l-diversity together with the risk of re-identification, differential privacy, which adds calibrated noise to protect individuals, federated learning, which trains models without centralizing raw data, homomorphic encryption, which allows computation on encrypted data, secure multi-party computation and trusted execution environments. Attacks such as membership inference that motivate these defences are discussed. A layered architecture for privacy-aware data management is proposed, with a research gap analysis, objectives and a comparison of techniques by protection level, performance cost and maturity. Regulatory drivers, including India's Digital Personal Data Protection Act 2023 and the GDPR, are considered. The review concludes that no single technique is sufficient and that layered combinations matched to the threat model provide the best balance of privacy, utility and performance.

DOI: https://doi.org/10.5281/zenodo.24032829

Published by: vikaspatanker

Authors: Gadge Anuja Harishchandra, Prof. N. G. Rokade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Deepfakes, realistic manipulations of faces, voices and videos produced with generative adversarial networks and diffusion models, threaten digital media integrity, personal privacy and public security. Traditional verification by visual inspection or simple forensic tools struggles against these sophisticated alterations. This paper reviews AI-based deepfake detection with emphasis on approaches suited to a real-time framework. It explains how deepfakes are generated and then surveys three families of detection cues: spatial artifacts within individual frames, such as warping, blending boundaries and texture inconsistencies captured by convolutional networks; temporal inconsistencies across frames, including unnatural blinking and lip motion detected with recurrent and spatio-temporal models; and biological signals such as subtle colour changes caused by blood flow that synthetic faces fail to reproduce. Benchmark datasets including FaceForensics++, Celeb-DF and the DeepFake Detection Challenge are discussed. A five-layer real-time detection architecture covering media ingestion, face processing, multi-cue analysis, decision fusion and reporting is proposed, together with research gaps, objectives, a comparison of methods and challenges such as generalization to unseen generators, compression and adversarial attacks. The review concludes that fusing complementary cues offers the most reliable automated defence.

DOI: https://doi.org/10.5281/zenodo.24030535

Published by: vikaspatanker

Authors: Maruboina Bhavitha, Manthena Chandra Prabhas, Mangali Namdev, Mamidi Pradeep Kumar, Umesh Kumar Kushva

Abstract: Lost and found management in college campuses and public spaces is frequently handled through informal channels such as WhatsApp groups, social media posts, notice boards and word of mouth. Although these channels support quick communication, they do not provide a persistent, searchable or accountable record of reports. Important information can become buried, duplicate reports can appear, and a person who has found an item may have no systematic way to connect it with the person who lost it. Ownership verification is also a concern because a photograph or general description alone is not always sufficient evidence of rightful ownership. FINDLY is a centralized web-based Lost & Found Portal proposed to organize the complete recovery lifecycle. The system follows the workflow REGISTER → REPORT → DISCOVER → MATCH → NOTIFY → VERIFY → RECOVER. Users can create lost or found reports containing category, description, location, date and photograph. Listings can be searched and filtered, while a multi-factor matching process can compare keywords, category, location, date and photo-related information to identify potential relationships. Notifications draw attention to possible matches, while a claim and ownership-verification stage separate matching from the final handover decision. Administrative moderation supports approval, rejection, duplicate flagging and dispute handling. Recovered items are marked as resolved so that the recovery lifecycle has an explicit endpoint. The project focuses on organizing fragmented information rather than relying on a single automated algorithm. The supplied project documentation provides the functional design, data model, implementation plan and test plan for the proposed system. Quantitative metrics presented in this paper are explicitly illustrative prototype values and should be replaced with measured project results before final submission.

DOI:

Published by: vikaspatanker

Authors: Prajakta Salve, Abhay More, M. Rayyan Ansari, Rushikesh Kande4

Abstract: Stress is a state that a lot of people experience while working on various activities at school or at work. As far as I noticed, stress begins in the brain. Moreover, stress manifests itself physically in some cases. Heart rate, heart-rate variability, electrical skin response, respiration, and blood flow are just some examples of body reactions to stress. Such manifestations may help to create stress detection technologies. However, only one body sign cannot serve as the only basis for stress assessment since movement, sitting position, temperature of the body, disease, medication, sensor placement and more may influence results. In my attempt to integrate the body monitoring with meditation, I would like to focus on relaxation through the use of technology. At the moment, I have the prototype of the system that operates using the benchtop. The system uses an ESP32, optical sensor, and an SH1106 OLED display. The optical sensor captures data related to the light. This data is processed using the MAXIM algorithm to calculate heart rate and blood oxygen levels (SpO₂). When the readings are valid the screen shows heart rate, SpO₂ and a rough stress level. This stress level is based on fixed thresholds. It is not an assessment.I have also started using the WESAD dataset. This dataset contains stress and emotional data from people. One subject’s file has been loaded into Python. The chest and wrist signals were checked. A short ECG segment was plotted. This step helps lay the groundwork for feature extraction and machine-learning model building. Now the system is still in stages. A machine-learning model has not been trained. A mobile app made with Flutter is not ready. Automated analysis before and, after meditation is not yet set up. The stress label used today is rule-based. It should not be treated as a measure of stress. This report goes over why the project was started what others have done in the field the hardware setup, how the software works what was done with the data, where things stand now what the limits are and what steps come next. The prototype is meant for teaching and wellness research. It is not meant for diagnosis. The prototype is the embodiment of the easy-to-use and non-invasive method of heart rate and SpO2 measurement, and stress estimation using some pre-defined rules. The application of machine learning methods for further stress classification was also considered within this study. Even though the suggested prototype is designed only for initial monitoring and education purposes, not for clinical diagnostics, it opens up possibilities for development of cheap technological solutions for stress monitoring and mental wellness.

DOI: https://doi.org/10.5281/zenodo.23977133

Published by: vikaspatanker