Automated Brain Tumor Segmentation Using a UNet3D-Based Deep Learning Model
A crucial task in medical imaging is brain tumor segmentation, which allows for accurate diagnosis and treatment planning for patients with brain tumors. Magnetic Resonance Imaging (MRI) provides detailed volumetric data, but manual segmentation is time-consuming and prone to variability. Deep learning, particularly convolutional neural networks such as UNet3D, has emerged as a powerful tool for automating and enhancing segmentation accuracy. Accurate and efficient segmentation of brain tumors from multi-modal MRI scans remains challenging due to the heterogeneity of tumor appearances, varying MRI modalities (e.g., T1, FLAIR), and the need for robust models that generalize across diverse datasets. This study aims to develop and evaluate a UNet3D-based deep learning model for automated brain tumor segmentation, leveraging the BraTS2020 dataset to achieve high-precision delineation of tumor regions in MRI scans. We developed and trained a UNet3D-based model tailored for brain tumor segmentation, utilizing PyTorch and nibabel to process 3D MRI data from the BraTS2020 dataset. The model was comprehensively evaluated on standard datasets, demonstrating robust performance across multiple MRI modalities. We conducted a thorough comparison with baseline segmentation techniques, including traditional methods and other deep learning approaches, analyzing metrics such as Dice scores and segmentation accuracy. Our results highlight the model’s superior ability to delineate tumor boundaries, offering improved precision and efficiency over baselines, thus advancing the application of artificial intelligence in medical imaging for brain tumor diagnosis.
Published by: Sidhartha Tadala, Angad Singh Chopra
Author: Sidhartha Tadala
Paper ID: V11I5-1178
Paper Status: published
Published: October 14, 2025
Fortifying AI Infrastructure: Securing Code, Configuration, and Integrity in National Systems
The rapid adoption of artificial intelligence (AI) on cloud platforms, such as AWS and Azure, has introduced critical security vulnerabilities across various national sectors, including defense, healthcare, and energy. While these environments deliver scalable intelligence, they also expand the attack surface, exposing misconfigured resources, unverified code, and weak identity controls. Recent breaches, including Capital One’s AWS data exposure, Tesla’s compromised Kubernetes console, and Microsoft’s AI dataset leak, demonstrate how cloud-hosted AI pipelines can be weaponized through insecure defaults, leaked credentials, and permissive access roles. This study analyzes prominent security incidents alongside current research on cloud and AI threats to identify recurring weaknesses in configuration management, secret handling, and model integrity. The findings highlight how attackers exploit these gaps to steal data, engage in cryptojacking, and gain unauthorized access to AI models. To address these risks, the paper proposes a framework for fortifying AI infrastructure that emphasizes: (1) zero-trust identity and access management, (2) secure coding and model lifecycle practices, (3) automated configuration scanning, and (4) continuous policy enforcement. The results underscore that AI infrastructure should be treated as national critical infrastructure, warranting rigorous standards and proactive defense measures. Without systematic hardening, AI pipelines are high-value targets for cybercriminals and nation-state actors, posing a threat to public safety and national security.
Published by: Ifeoma Eleweke
Author: Ifeoma Eleweke
Paper ID: V11I5-1190
Paper Status: published
Published: October 14, 2025
Libraries of Chandannagar: A Cultural Study with Special Reference to Akshar Bandhu Granthaghar
Chandannagar — a town with deep colonial and cultural roots — hosts a constellation of libraries that have historically mediated knowledge, memory, and everyday cultural practices. This paper analyses the evolving social roles of Chandannagar’s libraries with special reference to Akshar Bandhu Granthaghar (est. 2025). Using archival study, field observation, and semi-structured interviews with library users and staff across seven representative institutions (Akshar Bandhu Granthaghar; Chandannagar Pustakagar; Institute de Chandannagore; Chandannagar College Library; Chandannagar Museum Library; Gondalpara Sammelan Town Library; and selected parish/town libraries), we examine how mission, physical presentation (including cover-based selection), oral practices (storytelling, recitation), memory work, and nature-based reading activities contribute to inclusive reading cultures. Findings identify (1) a shift from elitist/academic library functions to community-embedded, democratic reading practices; (2) Akshar Bandhu’s explicit mission to facilitate book-familiarity among marginal groups through cover-driven selection and oral dialogic methods; and (3) hybrid practices that blend archival memory with living oral traditions. The study argues that community-centred libraries like Akshar Bandhu serve as models for democratizing reading and proposes policy and programming recommendations for sustaining such inclusive library ecosystems. The manuscript is prepared to meet international journal standards in Library & Information Science / Cultural Studies.
Published by: Dr. Patit Paban Halder, Dr. Somnath Bandyopadhyay, Dr. Kunal Sen, Dr. Sanjay Mukherjee, Dr. Basabi Pal, Dr. Manjusha Tarafdar, Mr. Agnidyuti Halder, Mrs Kabita Halder, Ms. Avishikta Halder
Author: Dr. Patit Paban Halder
Paper ID: V11I5-1175
Paper Status: published
Published: October 13, 2025
The Case of Declining Rental Properties All Over the World: A Global Perspective; Renting vs Buying
This paper tries to examine the factors that influence the choice of housing tenure for individuals across the globe. Factors such as demographics, social, economic and policy perspectives all play a key role in shaping the choice of tenure for those individuals. The choice is dependent on other factors like the age, household structure, price-to-income and price-to-rent ratios of the person, housing allowances and other such key factors. This paper also takes into consideration which choice is more prevalent in different countries and the reason behind it. It also takes into consideration factors such as the level of urbanisation and migration present in the country, as well as the societal and cultural norms of the country. Global Comparisons show that there are many differences in the choice of tenure, with developed economies showing vast differences between the amount of renters and owners, while developing economies like India face many challenges, including affordability problems, a high number of empty houses and uneven distribution between the number of renters and owners. This paper also takes into account the role of the government and their policies pertaining to taxes and incentives, reduction in interest on mortgages and home loans, as well as fewer taxes for homeowners promote homeownership, while poor and strict rules and regulations often make renting a more inefficient choice and have a major impact on the housing market for renters. Historical trends of housing taken from 2 decades, 2005-2015, 2015-2025, indicate changing patterns in the choice of tenure by individuals influenced by various key factors such as increasing prices and interest rates, higher inflation and fluctuations in the housing market have caused a difference in the choice of tenure. This paper also tries to examine the price-to-rent ratios and their impact on housing, along with the impact of interest rates, showing how these key factors can shift a person's choice from owning to renting or vice versa. Lastly, the paper analyses long-term sustainability and future outlooks for the housing tenures, highlighting the importance of policies as well as paying attention to the problems related to housing, such as vacancy and enhancing the understanding of the choice of consumers across the globe to figure out what tenure choice is most suitable for the future, with reference to their countries.
Published by: Avyukt Govil
Author: Avyukt Govil
Paper ID: V11I5-1186
Paper Status: published
Published: October 11, 2025
Intrusion Detection in AWS Cloud Environments Using Machine Learning on Network Flow Data
AWS Cloud Environments support core workloads and services, but are exposed to malicious actions and unauthorized activities in the transmission of network flow data. The threats subject cloud infrastructures to different types of attacks, thus Intrusion Detection in AWS Cloud Environments ensures privacy, reliability, and availability. This study explores the use of Machine Learning for intrusion detection by analyzing traffic patterns in cloud systems. The CSE-CIC-IDS2018 dataset, containing realistic benign and attack traffic, was employed for model training and evaluation. After comprehensive preprocessing and analysis, five Machine Learning algorithms were implemented: Random Forest, Decision Tree, Ridge Classifier, Logistic Regression, and Linear Support Vector Classifier. Their performance was measured using accuracy, precision, recall, F1 score, ROC-AUC, and detection time. Results showed that Random Forest and Decision Tree achieved the highest accuracy at 100%, with the Decision Tree demonstrating superior efficiency by classifying all instances in 0.056 seconds. Ridge Classifier followed with an accuracy of 99.2%, while Logistic Regression achieved 98.8%. The Linear Support Vector Classifier recorded the lowest performance with 96.2% accuracy. This research confirms the effectiveness of Machine Learning for cloud security. The Decision Tree Classifier, combining flawless accuracy with the fastest detection speed, emerges as the most practical model for real-time intrusion detection in AWS environments.
Published by: Oduwunmi Odukoya, Mariam Adetoun Sanusi, Samuel Adenekan
Author: Oduwunmi Odukoya
Paper ID: V11I5-1145
Paper Status: published
Published: October 8, 2025
Ancient Indian Scripture Based Retrieval-Augmented Systems: A Comprehensive Analysis
This paper focuses on the development and systematic comparison of Retrieval-Augmented Generation (RAG) systems, retrieval-only systems and LLM models all trained on ancient Sanskrit Scriptures. This was done in order to analyse whether RAG systems improved faithfulness in answers to reflective questions, by storing two pertinent Sanskrit scriptures: the Itihasa (including the Mahabharata and Ramayana) and the Bhagavad Gita in a FAISS index, I developed the following: a basic retrieval system from the FAISS index, a prebuilt LLM model (Qwen 2.5-3B-Instruct), an RAG system with the LLM model Qwen 2.5-3B-Instruct and an RAG system with Gemini 2.5 Flash. After development, I evaluated the four models on a list of twenty questions pertaining to philosophy, interpersonal and intrapersonal understanding, and emotional well-being. I ranked each answer on a scale from 1 to 5 on relevance, helpfulness, clarity and faithfulness. All retrieval and RAG models scored a perfect 5 in the ‘faithfulness’ metric in contrast to the base LLM model, which scored a 4.3. Moreover, I discovered that the use of a weaker LLM model in an RAG system can lead to worse results in the ‘helpfulness’ and ‘clarity’ metrics when compared to a regular LLM model when the retrieved verses are low. Through the methods and results of my research, I showed that RAG systems are necessary to provide specific and faithful answers from ancient Sanskrit philosophy.
Published by: Pradhyumna Prakash
Author: Pradhyumna Prakash
Paper ID: V11I5-1176
Paper Status: published
Published: October 6, 2025
