Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Madhab Chandra Das has a total of 22 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    ScreenForensics: A Multimodal Framework for Forensic Classification in Mobile Screenshots

    by Tasnia Tabassum, Madhab Chandra Das, Sajid Ahmed, Qingzhong Liu

    Mobile devices have become an essential source of digital evidence, and investigators frequently examine thousands of application screenshots containing authentication records, financial transactions, communication activities, location information, educational content, and system settings. Manual analysis of these screenshots is time-consuming and prone to human error, motivating the need for automated forensic evidence classification. This paper presents ScreenForensics, a multimodal framework that integrates visual and textual information through late fusion. The framework evaluates six deep learning image backbones and eight OCR-based text classifiers and combines the selected image and text models using ten multimodal fusion strategies, including score-level voting, rule-based fusion, and meta-learning approaches. Experiments were conducted on 61,319 Android screenshots from the Rico dataset organized into seven investigator-oriented forensic evidence categories. The selected LightGBM meta-classifier achieved an accuracy of 75.08%, a Macro-F1 score of 70.50%, and a Macro-AUC of 91.35%, outperforming the best image-only and text-only models. An ablation study shows that OCR-derived text provides the strongest individual discriminative signal, while visual information contributes complementary evidence that consistently improves classification performance. An inference dashboard is also presented to provide interpretable multimodal predictions for previously unseen screenshots. The proposed ScreenForensics framework provides an effective and practical approach for automated forensic screenshot classification. The framework is evaluated on a public mobile user interface dataset, while validation on real-world forensic case data remains an important direction for future work.

  • Open Access

    Explaining Cross-Dataset False Positives in Encrypted Traffic Classification for Forensic Interpretation

    by Madhab Chandra Das, Qingzhong Liu

    Machine learning models are increasingly used for encrypted traffic analysis because payload inspection is limited. However, high internal detection accuracy does not necessarily indicate that model outputs are reliable for forensic interpretation. This study examines this issue through cross-dataset false positive analysis. Four machine learning models are trained and evaluated internally on the CIRA-CIC-DoHBrw-2020 dataset. Random Forest achieved the highest internal accuracy of 99.62\%, while CatBoost achieved 98.92\% and was used for detailed cross-dataset and explainable analysis. The trained CatBoost model was then applied without retraining to 947,997 benign flows from the CESNET-TLS-Year22 dataset, where 41,238 flows were classified as malicious, resulting in a false positive rate of 4.35\%. This result shows that strong internal performance can still produce misleading alerts when the traffic environment changes. SHAP and permutation importance are used to examine feature dependence in the model predictions. Both methods identify a consistent set of aligned flow-level features, with packet size mean showing the strongest influence in the SHAP analysis. These features support classification but are not unique to malicious traffic and can also occur in benign encrypted communication. Therefore, feature importance alone is insufficient for forensic interpretation. The findings indicate that machine learning alerts should be treated as investigative leads rather than direct forensic evidence, and false positive analysis should be considered when interpreting model explanations across different traffic environments.

  • Open Access

    Machine Learning Prediction Is Not Enough for Forensic Interpretation: A Cross-Dataset Reliability Study of Network Traffic Classification

    by Madhab Chandra DAS, Sajid Ahmed, Qingzhong Liu

    High internal test performance does not always indicate that a machine learning model is reliable for forensic triage. This paper evaluates whether network traffic classifiers trained on CSE-CIC-IDS2018-AWS remain reliable when tested on external benign and malicious traffic. CESNET-TLS-Year22 is used to measure external benign false positives and true negatives, while USTC-TFC2016 is used to measure external malicious recall. All datasets are mapped into a common 18 feature representation. Four models are tested: Logistic Regression, Random Forest, XGBoost, and LightGBM. The results indicate an external reliability gap. LightGBM achieved the strongest internal performance, but it produced a 0.848025 false positive rate on CESNET benign traffic and detected only 0.006076 of USTC malware samples. Logistic Regression produced the lowest CESNET false positive rate, but it missed almost all USTC malware samples. SHAP analysis was used to inspect selected feature contribution patterns as decision support. These findings suggest that internal accuracy, F1 score, and ROC AUC should not be used alone for forensic decision support. External benign testing, external malicious recall testing, true negative review, threshold inspection, and explanation based review should be included before model outputs are used for forensic triage.

  • Open Access

    Backbone-Level Enhancements in YOLOv9 for Traffic Accident Detection from Video Footage

    by Sajid Ahmed, Tasnia Tabassum, Madhab Chandra Das, Uzair Hussain, Vung Pham

    Traffic accidents remain a major challenge for intelligent transportation systems, requiring reliable and real-time detection under complex visual conditions. This study aims to investigate how backbone-level architectural modifications affect traffic accident detection performance in video-based scenarios. A dataset of 250 accident videos was curated from a public traffic surveillance source. This resulted in approximately 3000 manually annotated frames covering diverse accident conditions such as motion blur, occlusion, and illumination variation. To improve detection performance, we introduce Cross Stage Partial (CSP)-based feature partitioning and extend Efficient Layer Aggregation Network (ELAN) structures within the YOLOv9 backbone. Experimental evaluation demonstrates that the CSP-enhanced YOLOv9-t model achieves the best performance among all tested variants, improving mAP50 from 0.35 to 0.50 (approximately 42.8% relative improvement) compared to the baseline YOLOv9-t model, while maintaining real-time inference speed. The results further reveal that CSP improves localization precision, whereas ELAN enhances recall, highlighting complementary behaviors of backbone-level modifications in traffic accident detection tasks. These findings provide insights into how targeted architectural refinements can improve detection robustness in challenging real-world traffic scenarios.

  • The application and comparison of Deep Learning models for the prediction of chest cancer prognosis

    by Mahadi Hasan, Miraz Al Mamun, Madhab Chandra Das, Maruf Hasan, Asm Mohaimenul Islam

    Lung cancers of all varieties, esophageal cancers, and cancers of the mediastinum (the area between the lungs), pleura (the membrane lining the chest cavity and surrounding the lungs), trachea, thymus gland, and heart are all classified as chest cancers, often known as thoracic cancers. Chest cancer can also spread from cancers that start in other places of the body. Chest pain is one of the usual signs of chest cancer, including hemoptysis or a cough that produces blood. Also, Coughing that hurts or a cough that does not go away is a sign of chest cancer. Mesothelioma, a cancer that begins in the lining of the chest or abdomen, frequently affects the lungs and other thoracic organs and tissues, which has prompted us to continue with this disease so that this research will aid in early detection. Chest X-rays and computed tomography (CT) pictures are the two diagnostic techniques that are most frequently utilized for these disorders. This study suggests a multiclassification deep learning model for detecting chest cancer using a dataset of chest CT-Scan pictures. While a chest CT scan is helpful even before symptoms show up and precisely detects the aberrant features that are found in images, a chest X-ray is less effective in the early stages of the disease.Furthermore, employing these kinds of photos will improve classification precision. To the best of our knowledge, no deep learning model in the literature can choose between these disorders. The current work considers the effectiveness of three architectures— CNN, ResNet50, and DenseNet121—. A thorough assessment of various deep learning architectures is performed using publicly available digital CT datasets with four classifications (Adenocarcinoma, Large Cell Carcinoma, Squamous Cell Carcinoma, and Normal). The study’s findings revealed that the DenseNet121 model performs better than the three other suggested models. CNN demonstrated 56.19% accuracy, whereas ResNet50 demonstrated 56.51% accuracy. The DenseNet121 model demonstrated 71.74% accuracy (ACC). We intend to investigate further deep learning models with large datasets.


  • A comparative study of machine learning approaches for heart stroke prediction

    by Madhab Chandra Das, Fatema Tabassum Liza, Partha Pratim Pandit, Fariha Tabassum, Miraz Al Mamun, Sharmistha Bhattacharjee, Shakil Bin Kashem

    The majority of strokes are triggered by the heart and brain blocking expected pathways. Today, it is the most common cause of death in the worldwide. By looking at the people affected, several risk elements that are thought to be connected to the stroke's cause have been determined. Numerous studies have been conducted for the prediction and categorization of stroke diseases using these risk variables. Similar to any diseases, an early diagnosis of a stroke can avert such occurrences and open the door to a healthy life. Machine learning (ML) techniques have been used in this study to accurately determine heart attacks. In order to determine multiple matrices like accuracy, recall, ROC, precision, and F1 score, we used nine different machine learning algorithms in this study, which include support vector machines (SVM), K-nearest neighbor (KNN), XGBoost, AdaBoost, Random Forest (RF), Decision Tree, LightGBM, and Logistic Regression. The results indicate that the Random Forest method outperformed the others with an accuracy of 98.4%.


  • SARS CovidAID: Automatic detection of SARS CoV-19 cases from CT scan images with pretrained transfer learning model (VGG19, RESNet50 and DenseNet169) architecture

    by Afia Farjana, Fatema Tabassum Liza, Miraz Al Mamun, Madhab Chandra Das, Maruf Hasan

    The COVID-19 outbreak has presented significant challenges to medical professionals worldwide and underscored the need for accurate and effective detection methods due to its highly contagious nature and potential for explosive community transmission. However, healthcare delivery has been hindered by a lack of testing kits. To address this, deep learning techniques have been utilized to diagnose COVID-19 using CT scans, which have higher sensitivity in detecting early pneumonic changes. However, limited access to large datasets of CT-scan images due to privacy concerns has made developing accurate models difficult. To overcome this, transfer-learning pre-trained models have been employed in this study to automatically detect COVID-19 cases from CT scan images. The proposed methodology utilizes VGG19, RESNet50, and DenseNet169 architectures to classify patients as COVID-19 (positive) or COVID-19 (negative), with DenseNet169 performing the best with an accuracy of 98.5% in predicting COVID-19 binary classification. The model showed no signs of overfitting or underfitting, with a great output curve relative to the training accuracy. The other models, ResNet-50 and VGG-19 showed performance well with an accuracy of 96.7% and 92.7%, respectively. However, VGG-19 had the lowest accuracy of 92.7%. The findings of this study demonstrate the potential of using machine learning methods for the accurate and timely prediction of COVID-19. DenseNet169 outperformed other models and provided better accuracy for the prediction of COVID-19 Binary Classification.


  • Machine Learning-Based Relative Performance Analysis for Breast Cancer Prediction

    by Fatema Tabassum Liza, Madhab Chandra Das, Partha Pratim Pandit, Afia Farjana, Asm Mohaimenul Islam, Fariha Tabassum

    The current high population growth in medical research has made early disease identification an urgent issue. The danger of dying from breast cancer is increasing exponentially along with the rapid population expansion. The second most serious cancer among those that have already been identified is breast cancer. Breast cancer poses a major hazard to women, as it is highly morbid and lethal. Doctors find it challenging to develop a treatment strategy that could increase patient survival time due to the lack of reliable prognostic models. As a result, developing a technique that results in the fewest errors while increasing accuracy takes time. A reliable, effective, and prompt response is offered by an automatic illness detection system, which also helps medical workers identify diseases and reduces the likelihood of fatalities. In this work, we investigate eight machine learning techniques, including GaussianNB, Decision Tree, K-Nearest Neighbor, Random Forest, support vector machine (SVM), XGBoost, LightGBM, AdaBoost. The Wisconsin Breast Cancer dataset was discovered in the UCI machine learning database, a well-known machine learning database. The performance of the study is assessed with relation to accuracy, precision, recall, F1 score, and ROC score. Among the eight-machine learning model, Random Forest and AdaBoost perform best which provides 99.20% accuracy and around 99% ROC curve score.


  • Heart Disease Detection Using ML

    by Ranjit Chandra Das, Madhab Chandra Das, Amzad Hossain, Ashiqur Rahman, Helal Hossen, Rakibul Hasan

    Hearth disease is one of the leading causes of death globally and a common disease in the middle and old ages. Among all heart diseases, heart attack and strokes are the most common cardiac illness that is the responsible majority of heart disease death. To identify heart diseases, for instance, Angiography is costly and has significant side effects. Therefore, machine learning can play an important role in identifying and predicting the potential risk factor of cardiac disease based on clinical and patient data, which is affordable and reliable. This study proposed and evaluated six machine learning models using survey data of 400k US residents to predict heart disease. This study also compared the evaluated six machine learning models, which are Xgboost, Bagging, Random Forest, Decision Tree, K-Nearest Neighbor, and Naïve Bayes. The accuracy, sensitivity, F1-score, and AUC of six machine learning algorithms are also evaluated and presented. In terms of performance results, the Xgboost model showed optimized results with an accuracy rate of 91.30%.


  • Predicting Chronic Kidney Disease Using Machine Learning Algorithms

    by Afia Farjana, Fatema Tabassum Liza, Partha Pratim Pandit, Madhab Chandra Das, Mahadi Hasan, Fariha Tabassum, Helal Hossen

    In the modern era, everyone tries to be aware of their health, but because of their workload and hectic schedules, they only pay attention to it when certain symptoms appear. However, because CKD (Chronic Kidney disease) is a disease with no symptoms or, in some cases, no symptoms at all, it is difficult to predict, detect, and prevent such a disease, which could result in long-term health damage. However, machine learning (ML) offers hope in this situation because it excels at prediction and analysis. In this paper, we proposed nine ML approaches, such as K-nearest neighbors (KNN), support vector machines (SVM), logistic regression (LR), Naïve Bayes, Extra tree classifiers, AdaBoost, Xgboost, and LightGBM. These predictive models are built using a dataset on chronic kidney disease with 14 attributes and 400 records to choose the best classifier for predicting chronic kidney illness. The dataset was gathered via Kaggle.com. Additionally, this study has compared how well these model's function. With the LightGBM model, we could predict kidney illness more accurately than ever before, with a 99.00% accuracy level.