Proceedings of International Conference on Network Security and Blockchain Technology

This paper focuses on a new algorithm of LSB embedding technique without Multiplier and Divider considering 8-bit binary bitwise operation for each RGB channels with its software along with hardware implementation. For the software implementation the performance quality metrics considered are PSNR and MSE. The result of both MSE and PSNR as compared with the recent research articles shows improvement of 92.78% in MSE and 1.66% in PSNR. In hardware implementation the prototype design using Xilinx–AMD for xc7vx485tffg1157-1 device FPGA is considered where area (top level module has 18 cells, numbers of I/O utilization (2% of total resources) and power (2.307 W) maintaining the security aspect of the proposed work for LSB embedding.

A combination of sensor nodes with limited storage, low data processing speed, and limited computing resources is known as wireless sensor networks (WSNs). Internet of things applications widely use sensors due to their low cost. The WSN environment still faces challenges in addressing security and privacy. Recently, Kar et al. presented an aggregate signature scheme that uses a certificateless primitive and has low computational costs. The security of the proposed CL-ASS lies in the hardness of solving the computational Diffie–Hellman problem (CDHP). Kar et al. stated that the CL-ASS is safe in a random oracle model. However, we assert that the scheme lacks security against two distinct kinds of attacks. We mount Type $$\mathcal{I}\mathcal{I}$$ I I attack and a coalition attack on the CL-ASS. We introduce improvements to the scheme, enhancing its efficiency, security, and conditional privacy-preserving capabilities by substituting dummy identities for the real identities of each sensor node.

Steganography is an approach that fabricates a confidential data within a cover media to obtain an stego-media without causing any noticeable differences. In comparison to text, audio, and video, image steganography becomes more prevalent due to social media, high invisibility, and enormous payload. This paper introduces an image steganographic technique that uses Difference of Gaussian (DoG) edge detection and its dilated version to classify edge pixels into three classes: non-edge, dilated-DoG edge, and DoG edge. In order to hide the secret bits within the cover pixels with X:Y:Z ratio, where, for any X, Y, and Z, X < Y < Z, the encoder employs all three classes. Experimental results indicate an average payload of 2.45 bpp with an average PSNR of 39.08 dB and an average SSIM of 0.986, demonstrating the method's effectiveness over existing approaches. Additionally, SR-Net steganalysis confirms the method’s security and ability to remain undetectable.

In the domain of medical imaging, maintaining the accuracy and safety of data is crucial. The paper initiates a novel medical image watermarking scheme to secure elevated imperceptibility and flexibility against various attacks. The technique combines the Local Binary Pattern (LBP) and Integer Wavelet Transform (IWT) to enhance authentication and protect patients’ privacy. The image is primarily decomposed using Integer Wavelet Transform (IWT), resulting in four subbands: (LL), (LH), (HL), and (HH), containing most of the image’s information, and is chosen to embed a watermark information. Before embedding, the watermark information is encrypted using LBP and a shared secret key $$\kappa $$ κ and stored into ( $$\gamma $$ γ ). The encrypted PG and LG are embedded within the HH subband to improve the watermarking process’s robustness and security. The inverse process is applied for watermark extraction: the watermark bits are extracted from the LL subband, and the metadata is retrieved from the LH subband. The HL is extracted from the HL subband, and the $$\gamma $$ γ is extracted from the HH subband, enabling verification of the image’s authenticity and integrity. Furthermore, a comparative analysis with previous watermarking techniques is conducted, demonstrating the superiority of the proposed method. Additionally, the temper of each band is detected to identify any potential alterations or tampering, ensuring the reliability of the watermarking process.

The benefits of aggregate signatures and certificateless cryptography are combined in certificateless aggregate signature (CLAS) methods. In particular, it reduces communication and computation costs by combining several signatures into a single aggregate signature and streamlining certificate handling without creating the key escrow issue. Although CLAS is an effective cryptographic tool, its security should be carefully examined prior to use. Here in this work, we draw attention to the insecurity of the Xu et al.’s scheme against fully chosen key attack. Here, we introduce the idea of equivalent validity and demonstrate that the proposed system does not fulfill the condition of equivalent validity since it is unable to survive fully chosen key attacks. Our cryptanalysis explains the fully chosen key attack over the CLAS technique defined by Xu et al.

In today's rapidly evolving digital landscape, robust cybersecurity measures are imperative to protecting digital assets and data integrity against constantly emerging threats, particularly malware. The primary objective of the proposed work is to establish a proactive malware detection system utilizing custom YARA rules tailored to identify distinct malware characteristics and patterns. The key contribution involves enhancing the system's ability to detect specific types of spyware, particularly custom PDF spyware, by implementing specialized rules and patterns. The yaraGenerator, yarGen, and yabin tools are used to check validation of proposed YARA rules. The yarGen tool yielded comparatively better detection results. After adding the import hashing technique, the generated YARA rules using the three chosen tools, yarGen, yaraGenerator, and yabin, are improved, and their efficacy on Advanced Persistent Threats, Ransomware, Malware Strings, Miner Strings, and Stealer is reassessed. This detection mechanism is prompting the development of unique methods for scrutinizing PDF files for malicious elements such as concealed scripts, deceptive content, and documented vulnerabilities.

One of the major drawbacks of conventional biometric system is that an attacker can steal or hack template and gain access to the system. Even though it is used for security purpose, biometrics itself is prone to attack. Template security in biometrics system is one of the thrust research areas. In this work, fingerprint system is used to encrypt templates. Performance is evaluated on PolyU DBI fingerprint database and MEPCO biometrics database. Fuzzy vault method is used to encrypt the template. Accuracy is 97 when keysize is 7 and 200 chaff points are concealed with genuine minutiae points. As security increased, accuracy decreases, as there is trade-off.

The present system is crafted to retrieve text from PDF documents and transform it into audio in various languages. It employs open-source resources, utilizing a web framework to handle text retrieval, translation, and audio synthesis. Users can submit a PDF, have the text pulled out, translated, and rendered into speech, offering accessible audio formats of documents, especially for those with visual disabilities. The system allows users to listen to text in their preferred language and aspires to enhance translation accuracy, broaden language selections, and integrate OCR for scanned files in the future.

In digital image watermarking digital media is safeguarded from illegal access, unauthorized modification, and copyright infringement. In this study, a watermarking scheme is represented by integrating the Gray Wolf Optimization (GWO) and Genetic Algorithm (GA) for intensifying the watermark embedding process. Here GWO and GA are hybridized to combine the strengths of both algorithms, where GWO provides a strong global search capability, while GA enhances the local search refinement, leading to improved optimization in the watermark embedding procedure. The Least Significant Bit (LSB) and IWT (integer wavelet transform) methods are employed for embedding the watermark, as they minimally affect the visual quality of the host image while maintaining simplicity and computational efficiency. By using the hybrid GWO-GA algorithm, the proposed approach optimally selects embedding positions in the host image’s pixel values, thus improving the robustness and imperceptibility of the watermark. Experimental results demonstrate that the hybrid optimization approach achieves high resistance against common image processing attacks while preserving the integrity of the host image. This hybrid model shows promise in striking a balance between watermark robustness and image quality, paving the way for enhanced digital rights management in multimedia applications.

With the rapid advancement of the Internet of Vehicles (IOV) and the increasing use of 5G networks, ensuring secure and efficient communication between vehicles and infrastructure has become crucial. This paper introduces the design and evaluation of a novel Authentication and Key Agreement (AKA) protocol specifically designed for IOV systems, taking advantage of the strengths of 5G technology. It incorporates advanced cryptographic methods, such as public key infrastructure (PKI), to establish secure communication pathways between vehicles, roadside units (RSUs), and other IOV components. In addition to improving security, the protocol emphasizes efficiency by optimizing message exchanges, lowering computational demands, and reducing communication delays. Formal verification using the AVISPA tool demonstrates that the proposed AKA protocol greatly enhances both the security and efficiency of IOV communications within 5G networks.

With the advancement of Information Technology, cybersecurity is in a continuous state of assessment as cyberthreats are increasingly leading to dangerous risks throughout the world. Cybercriminals often intend to steal, damage, or disrupt digital assets through a wide variety of malicious actions. Among these activities, phishing attacks are specifically conventional, targeting sensitive and valuable information including credit card numbers and passwords through fraudulent tactics. These attacks often involve fake emails, messages, or websites that imitate authentic sources to tempt individuals into revealing their confidential data or clicking on malicious links. To mitigate this growing threat, our paper proposes a nature-inspired algorithm-based solution for detecting phishing URLs, which offers an enhancement to the traditional dependency on multiple machine learning classifiers. By diligently extracting several features, a new phishing detection model has been developed, facilitating the ability to identify and combat these phishing attempts effectively.

Phishing has become a widespread issue in daily life, with many people falling victim to traps set by unethical actors, often resulting in the loss of money, private information, and more. The rise of deceptive phishing websites has prompted extensive research into automated detection techniques to safeguard users from these threats. This survey examines the application of machine learning algorithms in phishing detection, offering a comprehensive review of state-of-the-art models and techniques. By analyzing various detection approaches, feature selection methods, and datasets from existing studies, we explore how large-scale data is harnessed to train and enhance machine learning models. The objective is to highlight advancements in detecting phishing websites with greater accuracy and adaptability, thereby contributing to a more resilient framework for online security. Through this survey, we identify key challenges and potential future directions in phishing detection using machine learning, underscoring the role of algorithmic innovation in combating online fraud.

Suricata ( https://suricata.io/ ) is an open-source Threat Detection Engine owned and maintained by Open Information Security Foundation (OISF) ( https://oisf.net/ ). It works as an IDS and as well as an IPS to help cybersecurity analysts detect and protect against malicious network activities. Signature based rulesets play an important role in identifying and detecting malicious threats in a live network. Emerging Threats Open Ruleset ( https://rules.emergingthreats.net/ ) for Suricata detection engine can be found in /etc./suricata/rules directory of the Suricata distribution. These open rulesets for Suricata distribution are often referred as Suricata Rules. The open version of Suricata Rules consist of more than 36,000 rule signatures. These 36,000 signatures are not only just part of Suricata engine, but also are in active use in many other detection engines. In this paper, we will expose many interesting artifacts about Suricata Rules through frequency analysis. The insights gained from our analysis can be exploited to have a performance-optimized implementation of a threat detection engine using an open version of Suricata rules.

We showcase a critical security vulnerability in Machine Learning (ML) by introducing a novel synonym-based attack algorithm for generating adversarial examples from text data. While ML models are widely used in applications such as chatbots and spam detectors, there is limited research on their susceptibility to adversarial attacks, especially in text classification. Our algorithm creates adversarial examples by substituting words with synonyms while maintaining semantic integrity. We evaluate the algorithm’s effectiveness on three widely used text classification models—deep neural network, convolutional neural network, and long short-term memory—using the IMDB movie review and Reuters news datasets. The impact of the adversarial examples on model accuracy is assessed, with a particular focus on the threshold parameter M, which dictates the extent of text modifications. The results demonstrate that our algorithm generates adversarial examples that remain undetectable by human reviewers and spell checkers. When tested with a threshold M of 12, model accuracy declines significantly, ranging from 38.43 to 44.16%. Accuracy deterioration intensifies with increasing threshold values. Our research highlights a significant security gap in text-based ML systems and emphasizes the urgent need for robust defenses against proposed adversarial attack.

For a long time now, computers and humans have been separated using fully automated public turing tests, or CAPTCHAs. There are several variations of CAPTCHAs, including text-, picture-, audio-, video-, and math-based ones. To remain current, the CAPTCHA tests have undergone continuous updating. In order to work around the limitations of most traditional techniques, the stated model employs a flexible CAPTCHA creation process. The sophistication of bot traffic is rising, and conventional CAPTCHA techniques are finding it difficult to keep users satisfied while offering sufficient protection. In the first tier of the proposed model, there is an interactive Scratch Card CAPTCHA that engages users with an entertaining and user-friendly interface. A CAPTCHA for image recognition follows, which is meant to test users’ ability to complete visual tasks that are more challenging for machines to understand. In order to improve detection accuracy, the system makes use of machine learning models such as SVM (support vector machine), CNNs (convolutional neural networks), and RNN (recurrent neural network) that examine patterns of user engagement. This enables ongoing learning and modification depending on information gathered. Furthermore, concepts based on user experience are integrated to guarantee that the CAPTCHA procedure is still easy to use and captivating. Higher security metrics, better user interface, and experience with higher bot detection rates are anticipated results. By successfully thwarting illegal access attempts, this novel method not only seeks to make the Internet a safer place, but also guarantees a smooth and pleasurable experience for authorized users. The continued development and refining of this technology underline its potential to dramatically boost online security in an era of sophisticated automated attacks. In terms of security and usability, our method performs better than conventional CAPTCHA systems. With a 94.5% accuracy rate, the CNN model has considerable potential, and more refinement may increase the model’s resistance to hostile attacks.

This paper presents the development of Open-Audit, a blockchain-based compliance license management system designed to enhance transparency, efficiency, and integrity, creating a risk-free environment for all users. By leveraging blockchain technology, Open-Audit ensures accountability, while the integration of zero-knowledge proofs strengthens both security and privacy. The system involves three key stakeholders: regulatory bodies, software firms, and end users. Regulatory bodies oversee the compliance process, software firms request licenses, and end users can validate and provide feedback on the software. Smart contracts are deployed to manage user roles, license requests, comments, and approvals. Additionally, the InterPlanetary File System is used to store data securely in a distributed manner. With the incorporation of zero-knowledge proofs, Open-Audit ensures that sensitive information remains confidential while still providing verifiable proof of compliance. This innovative approach fosters a trustworthy ecosystem for license management, making open-audit both novel and effective.

Sudden epidemic infections have the potential of disruption of lives and economy. COVID-19 has provided a recent testimony to this. To prevent spreading of infection, contact tracing is mandatory. However, the contact tracing should be privacy preserving. In a smart city where inhabitants are empowered with body area network (BAN), this contact tracing can be done efficiently by examining the inter-BAN communications (IBC). To avoid tampering of IBCs, blockchain network maintains the IBCs. Contacts can be modeled as a graph. Existing works discuss the use of adjacency matrix for contact tracing in the graph. However, the contact graph may not be a dense graph and can have components in it. As a result, the time taken for determining the contacts of an user can be reduced if the contact tracing is confined within the component (or cluster) to which the user belongs to. The present work proposes a privacy-preserving contact tracing that determines the cluster to which the user belongs to and then finds all its direct and indirect contacts. Security of the proposed scheme has been minutely examined against the possible attacks.

Innovating mainstream technologies such as the Internet of Things (IoT) and Blockchain Technologies (BCT) are witnessed as technological breakthroughs with the potential to revolutionize our society. IoT is envisaged as a global network infrastructure that has outstretched the communication and information sharing not to just between typical computers and humans, but also numerous daily usable dumb objects. Although IoT's benefits are limitless, it has to face various scalability and security challenges for smooth adoption in real-world implementations. This also encumbers the proliferation of IoT applications in a large scale. However, these two technologies include technically dissimilar philosophies. BCT competence can be used as an integral part of resolving many shortcomings of IoT applications. This research work incorporates these two diverse technologies where BCT is employed in the context of IoT to provide a secure decentralized architecture to resist vulnerabilities and privacy threats.

The rapid expansion of Internet of Things (IoT) devices has brought significant cybersecurity challenges, especially in the realm of anomaly detection, which is crucial to defending against a diverse and evolving range of attacks. Conventional methods struggle to adapt to the dynamic and high-dimensional nature of IoT traffic, underscoring the need for advanced, data-driven solutions. This study presents an anomaly detection framework utilizing Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) to detect various types of attacks in IoT environments. Leveraging adversarial training, our method generates realistic data distributions, enabling the model to effectively distinguish between normal and anomalous traffic patterns. We trained and evaluated the WGAN-GP model on a comprehensive IoT dataset containing diverse attack types, including Denial of Service (DoS/DDoS), information gathering, man-in-the-middle, injection, and malware. A detailed correlation analysis helped identify feature dependencies, informing feature selection and dimensionality reduction. Experimental results demonstrate the model’s high accuracy (98.7%) and robust classification performance, with F1-scores consistently above 0.97 across attack categories. Additionally, the anomaly score distribution supports the model’s ability to detect anomalies with minimal false positives, showcasing its potential for dynamic thresholding. Our findings suggest that WGAN-GP is a promising approach for IoT anomaly detection, offering flexibility and scalability suitable for complex network environments. Future work could explore dynamic thresholding and further feature optimization to enhance detection accuracy and efficiency. This research introduces a scalable, adaptable framework for real-time IoT anomaly detection, with significant implications for improving the security and resilience of connected devices across diverse applications.

Amid rising urbanization and environmental challenges, sustainable solutions are increasingly vital. This paper examines the role of green roofs and vertical gardens in enhancing urban resilience by reducing air pollution. By integrating these green infrastructures with the Autoencoder-Decoder-based Long Short-Term Memory (AeD-LSTM) model, cities can significantly lower pollution levels and improve air quality, promoting healthier living environments. The research quantifies the pollution reduction potential of these green strategies, showing significant decreases in pollution levels post-implementation. It also highlights the synergistic advantages of combining green roofs and vertical gardens, contributing to more resilient urban ecosystems. The findings offer important insights for urban planners and policymakers advocating for green infrastructure to improve urban resilience and sustainability. The proposed model achieves 30 and 40% better accuracy in terms of $$R^{2}$$ R 2 -score and custom accuracy metrics, respectively. Also, this paper demonstrates a 30% energy savings using the energy consumption metric from the multi-LSTM autoencoder model.

The Internet of Things (IoT) has become popular across various applications. Still, it can also contribute to air pollution through increased energy consumption, electronic waste, and emissions from manufacturing and data centers. Enhanced logistics and traffic management can lead to more vehicle use, worsening urban air quality. This paper introduces a novel approach for predicting air quality in smart cities using the Fed-BiLSTM model, which stands for Federated Bidirectional Long Short-Term Memory. As urban areas adopt IoT technologies, accurate air quality monitoring and forecasting are essential for addressing environmental challenges. The proposed Fed-BiLSTM model leverages federated learning for secure, decentralized training across diverse data sources, ensuring privacy while enhancing prediction accuracy. Experimental results indicate a notable enhancement in the accuracy of AQI predictions—approximately 35%, 7%, and 10% better than traditional SVR, RFERF (ML models), and RNN (DL model) approaches, respectively. This work supports efforts to create sustainable urban environments by enabling informed decision-making through reliable air quality forecasts.

This work contributes toward the development of an efficient and scalable open-source Secure Multi-Party Computation (SMPC) protocol on machines with moderate computational resources. We use the ABY2.0 SMPC protocol implemented on the C++-based MOTION2NX framework for secure convolutional neural network (CNN) inference application with semi-honest security. Our list of contributions are as follows. Firstly, we enhance MOTION2NX by providing a tensorized version of several primitive functions including the Hadamard product, indicator function, and argmax function. Secondly, we adapt an existing Helper node algorithm, working in tandem with the ABY2.0 protocol, for efficient convolution computation to reduce execution time and RAM usage. Thirdly, we also present a novel splitting algorithm that divides the computations at each CNN layer into multiple configurable chunks. This novel splitting algorithm, providing significant reduction in RAM usage, is of independent interest and is applicable to general SMPC protocols.

This paper presents Min-Cart, an innovative framework designed to mitigate air pollution and reduce carbon footprints through an eco-friendly methodology. Effective strategies are crucial for sustainable development as urban areas grapple with escalating pollution levels and climate change challenges. Min-Cart integrates advanced data analytics and long short-term memory (LSTM) neural networks to reduce carbon footprints. Transportation optimization, industrial scheduling, and community engagement techniques are designed to assess air quality, forecast pollution trends, and minimization of carbon footprints, enabling timely interventions. By focusing on transportation optimization, industrial scheduling, and community engagement, the Min-Cart framework aims to foster greener practices. Experimental results demonstrate significant reductions in pollutant emissions and carbon output, highlighting the potential of eco-friendly approaches in creating healthier urban environments. Also, the Min-Cart framework achieves 58%, 32%, and 18% better prediction accuracies than SVM, Improved convnet and RNN, and CNN-ILSTM models. This work contributes to environmental sustainability and paves the way for future research in smart city initiatives and climate resilience.

Skin cancer is a rather widespread and potentially dangerous disease with aggressive malignancies that make fast metastasis in case of belated diagnosis. In turn, early detection of the disease significantly increases the chances of successful treatment and thus a favorable prognosis; therefore, diagnostic techniques need to be developed that possess both high sensitivity and specificity. The approach used in this paper employed the transfer learning skin cancer detection with advanced CNN techniques for classifying skin lesions as malignant or benign to enable early detection. The current study, therefore, focuses on the transfer learning methodology based on MobileNet architecture and has presented great potential regarding accuracy in malignancy detection. The next section presents a baseline comparison by proposing a custom CNN model for skin cancer classification, named DermCNN. Finally, we propose AccuDermCNN, a custom CNN model enhanced with edge detection techniques to further improve detection accuracy. Comparative analyses will demonstrate that all models, including AccuDermCNN, are on par with state-of-the-art methods and thus provide reliable diagnostic support. The results highlight that the integration of advanced CNN architectures, especially transfer learning and edge detection, can be effectively used for the medical diagnosis of skin cancer.

The paper presents a holistic Intrusion Detection System (IDS) that combines Machine Learning (ML) and Deep Learning (DL) to detect cyberthreats. The study aims at raising the accuracy of detection without false positives by combining the Host-Based IDS (HIDS) and Network-Based IDS (NIDS) frameworks. In addition, the study has experimented the various threat detection algorithms in dependence on logistic regression, k-nearest neighbors, naive Bayes, decision tree, random forest, and XGBoost, which can be applicable on the anomaly detection on top of the NSL-KDD dataset containing critical network activity attributes. It chose every model that was able to predict known threats by utilizing ML or DL for classifying complex patterns. Analysis suggests that the optimized random forest with Principal Component Analysis (PCA) will achieve the balance between accuracy and efficiency in computing. The observations made using XGBoost highlighted slight overfitting, which would require more fine-tuning. The artificial neural network balances depth and regularization to capture non-linear patterns without overfitting, using dropout and a deep architecture for optimal predictive performance in large feature spaces or intricate patterns. This study innovatively combines ML and a novel DL within the context of IDS, thus showing the way in which an integration of anomaly-based and signature-based detection will provide a robust and adaptive solution toward known and unknown threats within dynamic cyberenvironments.

Tumor classification becomes an essential requirement in the proper diagnosis and treatment approach of brain tumors. The present paper discusses an advanced implementation by combining the technique of UNet segmentation with Deep Maxout classifier for efficient tumor classification with precision. The convolutional neural network, UNet, would be utilized to perform a pixelwise segmentation that can lead to accurate identification of the region where the tumor is located within the MRI scan. The Deep Maxout classifier would then be implemented in order to classify the type of tumor with a high degree of accuracy. The proposed model trained and tested through publicly available brain tumor datasets was found to outperform state-of-the-art classifiers. It can be well compared with the known segmentation and classification techniques and can be applied to the clinic to aid in diagnosis and the development of personalized treatment plans for patients with a brain tumor.

This study's main goal is to perform a thorough analysis of the functional properties of Deep Learning models used with a particular dataset. The focus of this work is to carefully compare how well five different deep learning models perform in the task of categorizing photographs of gemstones from a dataset that contains eight different classes of gemstones. The focus is on resolving intrinsic dataset issues that come up during the model training phase. Five deep learning models were chosen for assessment: InceptionV3, ResNet50, MobileNetV2, and VGG16. These models were evaluated using the Gemstone Image collection. By doing a comprehensive analysis, we aim to understand the subtleties and capacities of every model and determine how well suited they are to deal with different situations that arise in Gemstone Image classification jobs. Five important performance indicators are included in our evaluation: F1-score, AUC-ROC score, recall, accuracy, and precision. We thoroughly examine each model's performance in terms of task classification, taking into account trade-offs between recall and precision as well as overall accuracy, prediction accuracy, and the capacity to identify pertinent cases. The models’ ability to discriminate across various thresholds is further elucidated by the AUC-ROC score. Our work attempts to clarify the benefits and drawbacks of these deep learning models by closely examining their performance across several evaluation criteria. This thorough comprehension will enable well-informed choices to be made about which deep learning models to use for gemstone picture classification tasks, hence resolving particular issues raised by the Gemstone Image dataset. The significant results obtained from our approach have led us to pursue a patent for the underlying methodology.

A multitude of researchers have commenced employing AI models in Automatic Image Captioning (AIC), particularly owing to progress in image analysis. These models integrate computer vision with Natural Language Processing (NLP) to connect visual data and human language by producing accurate descriptions of visual content. AIC surpasses mere labelling, facilitating a profound comprehension of images, essential for individuals with visual impairments. This technology improves the engagement and experience of visually impaired users by automatically delivering text-based descriptions, aiding their understanding of images encountered online and in their surroundings. The proposed AIC model is organised into a sequence of meticulous steps to guarantee robustness. It primarily emphasises data collection, integrating a varied assortment of images to enhance the model's efficacy. Subsequently, uncaptioned images are utilised to gather raw caption data. The model subsequently underscores the extraction of texture and appearance data from these images, which is essential for object identification and contextual understanding. This information is then employed for caption synthesis, wherein NLP techniques produce precise, contextually appropriate descriptions. The model is engineered to accommodate diverse scenarios and intricate visual contexts by improving its ability to generate pertinent target descriptions for each image, attaining an accuracy rate of 85%. This technology has potential applications in virtual assistants, image analysis, indexing and notably in assisting the visually impaired. AIC can greatly improve the daily experiences of individuals dependent on auditory descriptions by rendering visual content accessible. With the increasing interest in deep learning and natural language processing, automatic caption generation is anticipated to become more accurate and customised. The main aim of this project is to create a strong AIC model that successfully connects visual content with natural language, enhancing accessibility for individuals with visual impairments.

Domain adaptive deep learning framework designed particularly for the purpose of forest cover type deforestation in the Western Ghats, Tamil Nadu, India, using weakly supervised learning coupled with domain adaptation techniques applied to exploit the multi-temporal satellite imagery for understanding specific environmental issues in a biodiversity hotspot. It addresses the issue of variability in data caused by different ecological zones within the Western Ghats through the application of a DANN. The model will use the integration of U-Net for image segmentation and transformer-based models for sequential data processing. Preliminary results have shown considerable improvements regarding detection accuracy and generalization. This offers a real-time monitoring tool to support conservation efforts, thus working toward the achievement of SDG 13: Climate Action and SDG 15: Life on Land. The satellite imagery for this research is derived from advanced Earth observation platforms, including SPOT (Satellite Pour observation de la Terre) and WorldView series, which deliver high-resolution data appropriate for detailed environmental monitoring.

GeoLands presents a machine learning-based approach to assess landslide risks using high-resolution satellite imagery. By analyzing various geological and environmental features, the system predicts potential landslide-prone zones, supporting timely risk mitigation. Leveraging advanced image processing and predictive modeling techniques, GeoLands offers an efficient and scalable solution for enhancing landslide risk assessment, particularly in vulnerable and remote areas.

This paper proposes a new method for the classification of brain tumor types on MRI scans with firefly optimization along with parallel CNN. As accurate diagnosis of the tumor type is greatly needed in managing patients, our methodology improves upon the basic CNNs by integrating firefly optimization for feature selection and parameter tuning. In order to test our approach, we applied it to a large-scale dataset of MRI scans and illustrated notable improvements in classification accuracy as well as computational efficiency over the existing methods. Our results suggest that such a framework not only speeds up the process of classification but also enhances robustness in predictions, thus giving clinicians a quality tool for the early detection and treatment planning of brain tumors.

Alzheimer disease is the most prevailing type in dementia that causes permanent damage in the neuron connections and builds up dead cells results in behavioral changes, memory loss, and cognitive decline. So, an early detection of Alzheimer helps to manage the disease progression and helps to slow down rate of progression. The CNN model has been developed with an augmented data by setting the classification task with pre-processed images based on magnetic resonance imaging. The trained CNN model shows the capability to be a stage-wise classifier of the MRI scanned images into categories such as very_mild_demented, mild_demented, and moderate_demented besides the non_demented category. Various optimizers such as SGD, Adam, RMS-Prop as well as Adagrad are analyzed and studied. Performance of each fine-tuned model is evaluated by calculating Precision, Recall, F1-score, and Accuracy. Moreover, our proposed CNN Model has achieved accuracy of 99.69%, outperformed the existing models.

Air pollution poses significant threats to public health and the environment, necessitating effective prediction and control strategies. This paper presents an innovative model that combines K-Nearest Neighbors (KNN) and Artificial Neural Networks (ANN) to enhance the energy efficiency of air pollution prediction and control systems. The KNN algorithm is a robust tool for identifying similar historical data points, facilitating accurate forecasting of pollution levels based on various environmental factors. Complementing this, the ANN captures complex nonlinear relationships within the data, allowing for improved decision-making in pollution control measures. Our approach leverages a dataset comprising real-time pollution metrics, meteorological data, and energy consumption statistics. The model accurately predicts air quality indices through rigorous training and validation while optimizing energy usage. Evaluation results indicate that integrating KNN with ANN enhances predictive performance and promotes sustainable practices by minimizing energy expenditure in monitoring and controlling air pollution. This research contributes to developing smarter, more efficient environmental management systems, paving the way for healthier urban living environments.

Brain tumors are when cells in the brain grow abnormally and can be either non-cancerous or cancerous. These growths interfere with brain function and present symptoms such as headaches, seizures, memory loss, and cognitive challenges. Detection of them early is crucial for treatment. Current diagnostic techniques are time-consuming and heavily reliant upon radiologists. Errors in interpretation or delays in diagnosis could result in outcomes underscoring the importance of developing efficient diagnostic tools. With the advancements, in intelligence and deep learning technologies, today comes a rising demand for automated tools to aid in the detection of brain tumors efficiently and effectively. These systems can offer swift results to healthcare professionals for making informed decisions. The model presented in this context leverages learning methodologies like Convolutional Neural Networks (CNNs) to automate the identification of brain tumors from MRI scanned images. By analyzing a labeled dataset of MRI images and extracting features from them the model categorizes the scans into either being positive for a tumor or negative.