Thesis published in Dept. of CSE http://103.99.128.19:8080/xmlui/handle/123456789/40 2026-10-04T07:32:17Z 2026-10-04T07:32:17Z Anwar, Sayeda Suaiba ID:, 20MCSE019P http://103.99.128.19:8080/xmlui/handle/123456789/595 2026-10-04T06:17:57Z 2024-12-12T00:00:00Z

A Differential Privacy Aided Federated Learning Model for Intrusion Detection in IoT Networks Anwar, Sayeda Suaiba; ID:, 20MCSE019P In the rapidly-developing Internet of Things (IoT) ecosystem, safeguarding the privacy and accuracy of linked devices and networks is of utmost importance, with the challenge lying in effective implementation of Intrusion Detection Systems (IDS) on resource-constrained IoT devices. This study introduces a Differential Privacy-Aided Federated Learning architecture for intrusion detection in IoT contexts as a novel approach to addressing these difficulties. To build an intrusion detection model, we combined components of a Convolutional Neural Network (CNN) with Bidirectional Long Short-Term Memory (BiLSTM). We apply this approach to the Bot-IoT dataset, which was rigorously curated by the University of New South Wales (UNSW) and N-BaIoT dataset. Our major goal is to create a model that delivers high accuracy while protecting privacy, an often overlooked aspect of IoT security. Intrusion detection tasks are distributed across multiple IoT devices using federated learning principles to protect data privacy, incorporating the differential privacy framework to gauge and minimize information leakage, all while investigating the intricate relationship between privacy and accuracy in pursuit of an ideal compromise. We implemented two aggregation methods, namely, Federated Averaging (FedAvg) and Federated Proximal (FedProx). The trade-off between privacy preservation and model accuracy is investigated by adjusting the privacy loss and noise multiplier. Our research enhances IoT security by introducing a deep learning model for intrusion detection in IoT devices, explores the integration of differential privacy in federated learning framework for IoT and offers guidance on minimizing the accuracy- privacy trade-off based on specific privacy and security needs. Our study explores the privacy-accuracy trade-off by examining the effects of varying epsilon values on accuracy for various delta values for a range of clients between 5 to 25. We also investigate the influence of several noise multipliers on accuracy and find a consistent accuracy curve, especially around a noise multiplier value of about 0.5 for both the aggregation methods and with a higher accuracy for FedProx. The findings of this study have the possibilities to enhance IoT ecosystem security and privacy, contributing to the IoT landscape’s trustworthiness and sustainability. A Master of Science (M.Sc) Thesis in Computer Science and Engineering (CSE) Department at Chittagong University of Engineering and Technology (CUET).

2024-12-12T00:00:00Z Das, Avishek ID:, 21MCSE001P http://103.99.128.19:8080/xmlui/handle/123456789/587 2026-09-10T04:23:35Z 2025-06-02T00:00:00Z

AVaTER: A Multimodal Approach of Recognizing Emotion using Cross-modal Attention Technique Das, Avishek; ID:, 21MCSE001P Multimodal emotion classification involves the analysis and identification of human emo tions by integrating data from multiple sources, such as audio, video, and text. This approach leverages the complementary strengths of each modality to enhance the accu racy and robustness of emotion recognition systems. Audio data, for example, captures vocal tone and pitch, which are crucial for detecting emotions like anger or joy, while video data provides visual cues such as facial expressions and body language. Text data, often extracted from spoken words or written content, adds context and semantic depth to the emotion analysis. However, one significant challenge is effectively integrating these diverse data sources, each with unique characteristics and levels of noise. Additionally, the scarcity of large, annotated multimodal datasets in Bangla limits the training and evaluation of models. In this work, we introduced a novel multimodal Bangla dataset named MAViT-Bangla (Multimodal Audio Video Text Bangla dataset), which consists of 1002 samples incorporating audio, video, and text modalities. This dataset includes emotional categories such as anger, fear, joy, and sadness, providing a rich resource for emotion recognition studies in the Bangla language. Each sample in MAViT-Bangla was meticulously annotated to ensure high-quality labels, making it a valuable asset for researchers working in this domain. Moreover, we developed a framework for emotion recognition that utilizes a cross-modal attention mechanism among unimodal features. This mechanism facilitates the interaction and fusion of features from different modalities, enhancing the model’s ability to capture nuanced emotional cues. The proposed approach demonstrated its effectiveness by achieving an F1 score of 0.64, showcasing significant improvement over unimodal methods. This indicates that integrating multiple modalities through a cross-modal attention mechanism can substantially enhance the performance of emotion recognition systems, especially in the context of the Bangla language, where resources have historically been limited. The MAViT-Bangla dataset and our framework thus represent significant advancements in the field of multimodal emotion classification. A Master of Science (M.Sc) Thesis in Computer Science & Engineering (CSE) Department at Chittagong University of Engineering and Technology (CUET).

2025-06-02T00:00:00Z Hossain, G.M. Sakhawat ID:, 19MCSE005P http://103.99.128.19:8080/xmlui/handle/123456789/572 2026-09-06T05:39:19Z 2024-09-25T00:00:00Z

PDF Malware Detection Using Ensemble Learning Hossain, G.M. Sakhawat; ID:, 19MCSE005P Since Portable Document Format (PDF) is one of the most common file types, scammers put damaging code into people’s PDF files to get into their computers. Standard solutions and methods for identifying PDF malware are often insufficient to stop it completely. This is because PDF malware is very flexible and does not rely on a single set of traits. This work mainly aims to identify PDF malware quickly and effectively so that the problems can be addressed. To achieve the objective, initially various benchmark datasets were combined to generate an extensive dataset consisting of 15958 PDF samples, which considers the many types of behaviours exhibited by the samples, including non-malevolent, fraudulent, and deceptive behaviours. Three widely recognized PDF analysis tools (PDF-PARSER, PDFINFO, and PDFiD ) were employed to extract notable attributes from the PDF samples in the recently compiled merged dataset. Additionally, several types of derivations of traits were developed that have been shown through experimentation to help diagnose PDF malware. A technique was devised for constructing a highly effective and comprehensible set of characteristics by conducting a thorough empirical analysis of the retrieved and deduced features. Various standard machine learning models were examined which showed that the Random Forest classifier when using the final feature set, achieved an increase in accuracy of approximately 2%. In addition, the model’s explainability was showcased through the creation of a decision tree that produces rules that can be easily understood by humans. Finally, a comparative analysis was conducted with prior research that helped to highlight several significant findings. A Master of Engineering (M.Engg) Thesis in Computer Science and Engineering(CSE) Department at Chittagong University of Engineering and Technology (CUET).

2024-09-25T00:00:00Z Dey, Ashim http://103.99.128.19:8080/xmlui/handle/123456789/555 2026-09-06T04:05:34Z 2025-02-09T00:00:00Z

Hybrid Caching with a Dynamic Forwarding Strategy in ICN Dey, Ashim In this era of huge data traffic, the information-centric network (ICN) is a switch ing strategy from conventional host-centric communication to a new data-centric one. In-network caching an integral part of ICN can reduce the shortcomings of today’s location-based internet paradigm. As ICN has limited caching ca pability, the decision of optimal routing and content placement are the major challenges. In this study, we suggest a hybrid caching with a dynamic forward ing (HCDF) strategy which takes advantage of the off-path hash routing scheme and on-path caching scheme. To facilitate both caching schemes, a fraction of the caching space is allotted in each cache node for local caching. At the user-end segment, an estimation-based content placement (ECP) policy is designed using content popularity state and cache node position. This policy helps to cache pop ular content near the consumer’s location using the shared cache space. At the server-end segment, the HCDF scheme dynamically selects the most appropriate routing strategy among three strategies (symmetric, asymmetric, and multicast) based on hop count and path stretch. To demonstrate the effectiveness of HCDF, extensive simulation on real-world internet topologies has been carried out. From our investigation, it is observed that the suggested approach exhibits an impres sive performance in terms of cache hit, latency, and average link load compared to other standard hash routing and well-known on-path caching schemes.ICN; Hash routing; Hybrid caching; Content placement; Caching strategy A Master of Science (M.Sc) Thesis in Computer Science and Engineering (CSE) Department at Chittagong University of Engineering and Technology (CUET).

2025-02-09T00:00:00Z