Journal of Science Engineering Technology and Management Science

Bot cyberbullying has emerged as a critical issue in digital communication, with over 59% of teens reporting experiences of online harassment and more than 70% of such cases occurring on social media platforms. Recent studies show that multi-class bot cyberbullying datasets often contain up to 25,000 labeled instances spanning categories such as insult, threat, racism, and sexism, frequently exhibiting severe class imbalance and linguistic variation. Manual detection methods suffer from subjectivity, inconsistent labeling, and an inability to scale with the rapid influx of user-generated content. Conventional machine learning approaches are limited by shallow feature representation, low recall on minority classes, and poor detection of implicit or masked abuse. Additionally, existing research often overlooks the integration of deep ensemble learning methods with optimized preprocessing and contextual analysis. To address these limitations, this study proposes a hybrid Multiclass Unmasking Bot Classification system that integrates a novel combination of feature-rich N-gram extraction with dual deep learning classifiers: a Deep Neural Network (DNN) and a Convolutional Neural Network (CNN). The process begins with dataset ingestion and detailed Exploratory Data Analysis (EDA) to assess data distribution and class imbalance. Text preprocessing follows, including tokenization, lemmatization, and noise removal. The cleaned data is then vectorized using TF-IDF with bi-gram support to capture both isolated and contextual word associations. The DNN is employed to capture deep semantic hierarchies, while the CNN is used to identify local linguistic patterns. This parallel dual-stream architecture ensures robust learning across diverse types of bot-generated cyberbullying. Finally, the trained models are evaluated for prediction accuracy and class-wise performance, significantly outperforming baseline classifiers in terms of precision, recall, and F1-score.

Keywords : Cyberbullying Detection, Deep Learning, Multiclass Classification, TF-IDF with Ngrams, Convolutional Neural Network (CNN).

Author : K. Swayam Prabha, K. Murali Krishna, G. Ranjith Kumar, I. Laxman, D. Arun Kumar

Title : DEEP LEARNING FRAMEWORK FOR MULTICLASS BOT CYBERBULLYING DETECTION ON SOCIAL MEDIA

Volume/Issue : 2025;02(08)

Page No : 635-645

India is rapidly transitioning toward the electrification of public transportation—particularly electric buses—to reduce carbon emissions and dependency on fossil fuels. As the country's energy demands continue to grow, electric buses have emerged as a key solution for sustainable urban transport. By 2022, India had more than 4,000 electric buses in operation, a significant increase from just a few hundred in 2017. This growth reflects the nation's strong commitment to reducing pollution and improving energy efficiency. To further enhance operational performance, a data-driven approach using machine learning is being applied to predict energy consumption and optimize economic performance in electric city buses. This approach aims to improve the overall efficiency and sustainability of urban public transportation systems. Traditionally, electric bus management relied on static route planning, manual scheduling, and historical data, making the system reactive rather than proactive. Maintenance and energy optimization were often performed only after issues occurred, leading to inefficiencies and increased operational costs. These systems lacked predictive analysis, resulting in inefficient energy use, suboptimal route planning, and unpredictable expenses. As urban centers face rising energy demands and environmental pressures, there is a clear need for intelligent, dynamic fleet management. A machine learning-based system can analyze real-time data from electric buses—including battery levels, route patterns, and weather conditions—to accurately predict energy consumption and operational costs. This enables dynamic route optimization, predictive maintenance, and load balancing, significantly reducing energy waste and operational costs while enhancing the efficiency of public transport systems. Additionally, AI-driven models will provide real-time insights, allowing fleet operators to make proactive decisions that improve overall performance and reliability.

Keywords : Electric Buses, Machine Learning, Energy Consumption Prediction, Sustainable Urban Transport, Predictive Maintenance

Author : K. Madhavi, S. Rajesh, R. Maliklal Naik, B. Naveen

Title : SMART ENERGY FORECASTING FOR ELECTRIC CITY BUSES USING MACHINE LEARNING

Volume/Issue : 2025;02(08)

Page No : 646-657

Embedding secret data within digital images (steganography) is a widely used technique for covert communication. However, traditional LSB methods lack confidentiality and integrity checks, making them vulnerable to extraction and tampering. To address these shortcomings, this work proposes an integrated solution that combines encryption (ECC or ChaCha20), LSB steganography, and compression to securely hide short payloads inside cover images. Public-key cryptography offers strong security but can be computationally intensive, whereas symmetric stream ciphers like ChaCha20 deliver high performance but require key distribution. Conventional LSB steganography directly embeds plaintext bits, exposing messages to steganalysis and eavesdropping. A robust system is therefore needed to ensure that extracted data cannot be read without the key, and that any alteration is detectable. In this project, the secret message is first encrypted using either ECIES over secp256k1 (asymmetric) or ChaCha20 (symmetric), and its SHA-256 hash is computed to ensure integrity. The resulting ciphertext bytes are Base64-encoded and converted into an ASCII bitstream, which is then embedded into the least significant bits of an RGB image. The stego image is subsequently compressed using zlib to reduce file size and obscure potential embedding artifacts. On the receiver side, the compressed file is decompressed, the LSB bits are extracted to reconstruct the Base64 ciphertext, the SHA-256 hash is recalculated to verify integrity, and the payload is decrypted. A Tkinter GUI guides users through uploading images, entering secrets, selecting encryption modes, and viewing performance graphs that compare ECC and ChaCha20. Experimental results show that ChaCha20 encryption is significantly faster than ECC for short messages, while ECC provides asymmetric security without the need for key exchange. By integrating strong encryption, steganography, and compression, this system mitigates the limitations of traditional LSB methods— ensuring confidentiality, integrity, and minimal perceptual distortion—making it well-suited for applications requiring covert, tamper-evident communication

Keywords : Steganography, ChaCha20, Elliptic Curve Cryptography (ECC), Confidentiality and Integrity, LSB Embedding

Author : J. Prashanthi, P. Preethi, M. Vignesh, J. Tharun, P. Ravinder

Title : SECURE IMAGE STEGANOGRAPHY USING ECC AND CHACHA20 WITH COMPRESSION

Volume/Issue : 2025;02(08)

Page No : 658-668

Injection attacks in APIs are a significant security concern, where malicious actors exploit vulnerabilities by inserting harmful code into API requests. These attacks can lead to unauthorized access, data breaches, and system compromises, ultimately affecting the integrity and confidentiality of applications. Historically, the detection of injection attacks in APIs has evolved alongside the increasing reliance on APIs in modern software development. Early detection methods primarily involved manual code reviews and pattern matching, which were often insufficient due to the complexity and variety of injection techniques. As APIs became more prevalent, the demand for automated and sophisticated detection mechanisms grew. Traditional systems for mitigating injection attacks have relied on input validation, parameterized queries, and the use of Web Application Firewalls (WAFs). These methods aim to prevent malicious inputs from being processed by the system. However, they frequently fall short in detecting complex or novel attack patterns, leaving systems vulnerable. The motivation for developing advanced detection systems arises from the limitations of traditional approaches. The increasing sophistication of attackers and the critical role of APIs in modern applications necessitate more robust and intelligent security measures. This research is driven by the need to enhance detection capabilities, reduce false positives, and adapt to emerging threats in real time. A core problem with traditional systems is their reactive nature and limited adaptability. They often require manual updates to recognize new attack vectors and struggle to detect obfuscated or zero-day attacks. Additionally, their reliance on predefined rules results in high false positive rates, where legitimate traffic is mistakenly flagged as malicious. To address these challenges, the proposed system leverages Bidirectional Recurrent Neural Networks (BRNNs) to detect injection attacks in APIs. BRNNs process input sequences in both forward and backward directions, capturing contextual information more effectively than unidirectional models. This approach enables the system to identify complex patterns associated with injection attacks, thereby improving detection accuracy and adaptability to evolving threats.

Keywords : Injection Attacks, API Security, Bidirectional Recurrent Neural Networks (BRNNs), Detection Accuracy, Automated Threat Detection

Author : Dr. A. Yashwanth Reddy, Sri Harsha Dixit, SV. Swathi, N. Shashank, K. Prashanth

Title : AI DRIVEN DETECTION OF INJECTION ATTACKS IN API’S USING BIDIRECTIONAL RECURRENT NEURAL NETWORKS

Volume/Issue : 2025;02(08)

Page No : 669-680

As cloud computing continues to change, maintaining strong data security is still a major concern, especially in light of growing threats to encryption and key management systems. In order to overcome these vulnerabilities, this study proposes a novel, two-phase method that combines blockchain-based key management with dynamic Advanced Encryption Standard (AES) encryption. Our system uses dynamically generated AES keys for every file, in contrast to conventional static encryption techniques, reducing the possibility of widespread data compromise in the case of a key breach. In addition, blockchain technology is used to decentrally and impenetrably store encryption keys and other critical metadata, guaranteeing improved availability and integrity. To improve data security even more, Elliptic Curve Cryptography (ECC), a thin yet effective security layer, is implemented for safe key transmission and file exchange. The suggested methodology dramatically lowers attack surfaces in cloud environments by doing away with the need for centralized key storage and using dynamic encryption. Its scalable and modular architecture enables smooth interaction with current cloud infrastructures, providing businesses and people with an adaptable, future-ready solution for protecting sensitive data in the cloud.

Keywords :

Author : Umme Habeeba, Dr. Mohammad Pasha, Samreen Sultana

Title : ENHANCING CLOUD DATA SECURITY WITH DYNAMIC AES ENCRYPTION AND BLOCKCHAIN-BASED KEY MANAGEMENT

Volume/Issue : 2025;02(08)

Page No : 681-688

Stroke is one of the most prevalent causes of death and disability in the world, although it is preventable and treated. Improving clinical outcomes and lowering the burden of disease are significantly aided by early stroke detection and prompt treatments. Because machine learning techniques can be used to detect strokes, they have garnered a lot of attention in recent years. Finding trustworthy techniques, algorithms, and characteristics that support healthcare providers in making well-informed decisions on stroke prevention and treatment is the goal of this project. In order to accomplish this, we have created an early stroke detection system that uses CT scans of the brain and a ResNet (Residual Network) model to identify strokes at an extremely early stage. The ResNet model is used for image classification in order to obtain the most pertinent features for categorization. The system's efficacy was assessed using cross-validation using metrics including precision, recall, F1 score, ROC (Receiver Operating Characteristic Curve), and AUC (Area Under the Curve). The suggested diagnostic system enables doctors to treat stroke patients with knowledge.

Keywords :

Author : Shagufta Areej, Dr. I. Samuel Peter James

Title : A DEEP LEARNING-BASED DIAGNOSTIC MODEL USING NEUROIMAGES (BRAIN-STROKE DIAGNOSIS)

Volume/Issue : 2025;02(08)

Page No : 689-695

The health and well-being of people are enormously impacted by sleep problems, including insomnia, sleep apnoea, and other illnesses. The quality of life for those who are impacted can be improved by early diagnosis and successful treatment made possible by an accurate and effective classification of various conditions. For categorisation, the current systems mostly use Artificial Neural Networks (ANN), which are efficient but sometimes computationally demanding and difficult to understand. In order to categorise sleep disorders, this study suggests a Random Forest-based method using a dataset of 400 samples with 13 pertinent variables. The Random Forest model was chosen because it is robust, easy to understand, and has a higher capacity to manage intricate, non-linear interactions in the data. The study uses this algorithm to categorise sleep disorders into three groups: sleep apnoea, insomnia, and none. Its performance is better than that of conventional ANN-based systems. Standard performance criteria, such as accuracy, precision, recall, and F1-score, are used to evaluate the Random Forest model. The results demonstrate that the suggested method performs better than current models, providing improved accuracy and dependability in the categorisation of sleep disorders.

Keywords :

Author : Sana Fathima, Dr. Mohammad Pasha, Samreen Sultana

Title : ADVANCED MACHINE LEARNING MODELS FOR SLEEP DISORDER DETECTION AND CLASSIFICATION

Volume/Issue : 2025;02(08)

Page No : 696-704