RKRaihan Khan

Raihan Khan

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Raihan Khan has a total of 11 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    MCFN: A Lightweight Multi-Path Compound Fusion Network for Deepfake Face Detection

    by Susmita Saha Lagna, Md Raihan Khan

    The proliferation of deepfake media poses severe threats to digital trust, biometric security, and information integrity. Most existing detection methods rely on a single feature domain, limiting their ability to generalise across diverse manipulation techniques. In this paper, we propose the Multi-Path Compound Fusion Network (MCFN), a novel architecture that jointly exploits texture, edge/gradient, and frequency-domain cues through three parallel encoders. A Compound Fusion Module enriches the per-path embeddings via cross-path multi-head self-attention and compresses them into a single conditioning token, which dynamically modulates an EfficientNet-B0 backbone through Feature-wise Linear Modulation (FiLM). A primary classifier head and an auxiliary frequency forensic head are jointly optimised with a composite loss comprising label-smoothed binary cross-entropy, an auxiliary frequency loss, and a supervised contrastive loss. Extensive experiments on four benchmarks---FaceForensics++, Celeb-DF, the Google DFD benchmark, and a combined FF++\,+\,Celeb-DF corpus---show that MCFN attains the best accuracy on all four benchmarks and the best or comparable AUC relative to recent state-of-the-art methods, while maintaining only 6.33\,M trainable parameters. A systematic ablation study confirms the contribution of each component, with FiLM conditioning and cross-path attention yielding the largest individual gains. These results establish MCFN as an effective, lightweight, and generalisable framework for deepfake face detection.

  • Open Access

    Equivariant Geodesic Networks: Geometry Preserving Learning on Riemannian Mani-folds

    by Md Raihan Khan, Airin Akter Tania

    Many high-dimensional data modalities-including covariance descriptors, diffusion tensors, and kernel matrices-naturally reside on Riemannian manifolds such as the space of Symmetric Positive Definite (SPD) matrices. However, conventional deep neural networks often fail to respect the intrinsic geometry of such data, leading to suboptimal representations and generalization. We introduce Equivariant Geodesic Networks (EGN), a novel architecture designed to operate directly on Riemannian manifolds while preserving key geometric properties. EGN incorporates manifold-consistent operations, including equivariant mappings, adaptive geometric bias, and structured low-rank updates that respect the underlying topology. Unlike existing methods that either flatten or project SPD data into Euclidean space, EGN directly learns on the manifold, preserving geometric consistency throughout. We provide theoretical analysis of the manifoldpreserving properties of our layers and demonstrate significant empirical gains on tasks involving SPD-valued data, such as EEG-based emotion recognition and imagined speech classification. EGN outperforms existing Euclidean and pseudomanifold baselines, offering a principled approach to end-to-end learning on Riemannian data manifolds.

  • Open Access

    DAFNet: A dual-path attention fusion network for EEG emotion recognition via CNN and graph-based global modeling

    by Raihan Khan, Airin Akter Tania, Tanjum Arifen Bushra, Jahanara Pritha, Mohiuddin Ahmad


  • Custom Dataset-Driven Unsupervised Low-Light Image Enhancement Using 2D CNN

    by Airin Akter Tania, Raihan Khan, Mohiuddin Ahmad

    Low-light image augmentation, which seeks to improve image visibility and quality under low illumination, is an important job in computer vision. We investigate underexposure image improvement in this study using a variety of datasets, including three established datasets (LOL, ExDark, SICE) and our custom dataset, AR. By merging these datasets, we produced a unique dataset. We propose a novel method for training a robust model using both standard and custom datasets to optimize performance under poor lighting conditions. To guide model training, we analyze five important loss functions: loss of exposure, spatial consistency loss, color constancy loss, illumination smoothness loss, and total loss. These losses improve the image in low-light conditions while ensuring the retention of significant visual elements. spatial consistency loss Additionally, we have made our dataset publicly accessible to facilitate further research. According to experimental results, our model outperforms prior benchmarks and produces encouraging results in terms of improving image quality. Our approach and the AR dataset offer significant resources to the low-light image improvement community, establishing the groundwork for further developments.


  • Machine Learning Techniques for Brain Stroke Analysis and Prediction

    by Rabita Hasan, Sheikh Md. Rabiul Islam, Raihan Khan

    Brain strokes a major contributor to global mortality and morbidity requires prompt prediction and intervention to reduce its impact. The study explores the practice of machine learning techniques for predicting brain strokes with the goal of improving early detection and preventive strategies. Utilizing a comprehensive dataset that includes demographic, lifestyle, and medical factors this study employs various machine learning systems such as logistic regression, random forest, decision tree, K-nearest neighbors, AdaBoost, XGBoost, and voting classifiers to develop predictive models. Feature ranking techniques are applied to detect the most significant variables which causes the incidence of stroke. The analysis incorporates preprocessing steps including data normalization, cleaning and managing missing values to determine the quality of the input data. Performance metrics including F1-score, recall, accuracy and precision are applied to measure the proficiency of the established models. Decision tree, XGBoost and voting classifier are showing best performance in this study with a similar accuracy of 96%. Additionally, this study contributes to build up a stacking model which achieved higher performance with an accuracy of 97%.


  • Smart Classroom Automation: A Fusion of AI with Voice, Gesture, and Face Recognition Attendance System

    by Raihan Khan, Abdullah Al Ahad, Airin Akter Tania, Tithi Das, Bibhakar Das

    From the very beginning of the twenty-first century, every task has started to be digitalized and computerized. In this modern age, education, being the most powerful tool for transforming the world into a better place, has led to a high demand for smart classrooms. This paper introduces an automated system designed to transform traditional classrooms into intelligent learning environments. A smart switching system for regulating electrical load is proposed here. A gesture-based interface to ensure exact control over the sound system has been proposed, adding a layer of intuitive interaction. Additionally, this system simplifies attendance tracking by integrating real-time face recognition of students, facilitated by a camera positioned above the classroom. Promoting easy access and interaction, users can operate a variety of classroom electrical devices using natural language voice commands. The system has been tested under different environments, including variable lighting configurations, distance between person and camera, moving subjects, and a diversity of persons. The suggested system responds to electrical loads in 0.68 seconds after commanding. The automatic sound system has an accuracy of 94.12%. The facial recognition technology has a 98 % recognition accuracy. Experimental findings reveal the system's convenient operation, good safety, and strong practicability.


  • Mental Stress Detection from EEG Signals Using Comparative Analysis of Random Forest and Recurrent Neural Network

    by Raihan Khan, Mohiuddin Ahmad

    Mental stress is a common problem that affects people in numerous facts of their lives, and early discovery is critical for effective treatments. This study introduces a unique approach using sophisticated methods like Recurrent Neural Network (RNN), Random Forest, and Electroencephalogram (EEG) signal analysis. By analyzing EEG signals, the aim is to quickly and accurately identify signs of mental stress. The model is built using the GAMEEMO dataset and validated with the SEED dataset, focusing on recognizing patterns in EEG signals associated with varying stress levels. The model demonstrates high accuracy, precision, Fl score, and recall values, showcasing its effectiveness. With an 87 % accuracy for arousal and 83 % for valence, the RNN model outperformed Random Forest, which achieved 83% accuracy for arousal and 75% for valence. This study advances mental health monitoring and intervention systems, paving the way for practical application and better well-being.