Research Publications

Unlike earlier forms of disinformation that were primarily text-based, modern rumors often combine both text and images to construct compelling but deceptive stories. Detecting such multimodal rumors presents new challenges, as it requires understanding not just each piece of content in isolation, but how they work together to mislead. Traditional rumor detection models aim to identify such content by relying on textual cues or employing shallow multimodal fusion techniques to align modalities of information. However, these approaches often struggle to capture the complex semantic interplay between modalities, especially when faced with semantic discrepancies or emotionally enforced visual cues. This is due to their inability to effectively model the subtle semantic variation, hierarchical, and non-linear relationships that exist between different modalities. Despite recent advances, existing rumor detection models face two critical challenges: (1) Existing methods often treat semantic discrepancies or ambiguities as noise during the fusion of visual and textual modalities. This neglect causes the models to overlook essential cross-modal interdependencies, ultimately leading to poor performance. (2) Most approaches assume a one-to-one correspondence between visual and linguistic features, ignoring the richer hierarchical relationships that bridge different semantic levels across modalities. As a result, models often fail to capture the layered structure of meaning and overlook important hierarchical cues. To address these challenges, this study presents the following key contributions: (1) The study introduces RADAR, a framework that treats semantic discrepancies not as noise to be reduced or ignored, but as valuable signals that can enhance detection. RADAR focuses on amplifying them by incorporating three modules: First, the Semantic Discrepancy Amplifier (SDA), a multi-head attention mechanism that dynamically detects and amplifies mismatches between modalities. Second, these differences are improved using Progressive Feature Purification (PFP), which adjusts filters to remove unnecessary noise while keeping important signals. Third, the Feature Feedback Loop (FFL) continuously improves the quality of the features. Together, these components form a pipeline that effectively highlights subtle inconsistencies often found in rumor content. RADAR achieves state-of-the-art results, with 96.0\% accuracy on the Twitter dataset and 91.2\% on Weibo, effectively detecting both rumors and non-rumors by amplifying semantic discrepancies. (2) This paper proposes the RumorCone framework that incorporates semantic hierarchies using hyperbolic geometry to preserve hierarchical, non-linear relationships, rather than representing them at a flat semantic level. It decomposes both image and text content into three levels: base, mid, and high-level abstractions. First, it places these different types of information into a hyperbolic space to preserve their natural tree-like shape and show how they interact non-linearly. Second, the model uses cross-modal hyperbolic attention, which focuses on important features from each modality at every level while maintaining consistency across modalities. Additionally, hierarchical semantic processing ensures that text and image features are aligned at each level of abstraction, further enhancing the model's ability to identify misalignments. Third, it uses a flexible fusion method that dynamically adjusts the contribution of each modality based on its alignment. RumorCone achieved improved performance 86.76\% F1-score on MR²-E and 85.48\% on MR²-C by utilizing hierarchical reasoning in hyperbolic space. The study further evaluates through cross-lingual experiments, gathering 100 recent social media posts in five languages and including various scenarios across multiple platforms. It achieved 83\% accuracy on high-resource languages such as English and Chinese; performance declines to 68\% for Bengali and Hindi due to translation-induced semantic shifts, with baselines performing even worse. The research demonstrated that semantic discrepancies serve as significant indicators of rumor content, while hierarchical semantic reasoning enhances the accuracy of content-based detection. Cross-lingual evaluation revealed the necessity for datasets covering low-resource languages.

Md Mahbubur Rahman

Traditional rumor detection models attempt to identify such content by relying on textual cues or employing shallow multimodal fusion techniques. However, these methods often assume a simplistic one-to-one alignment between modalities, overlooking the richer hierarchical relationships across modalities, failing to capture the layered structure of meaning. In this paper, we present RumorCone, a novel method that employs hyperbolic geometry in order to preserve hierarchical, non-linear relationships, rather than representing them at a flat semantic level. First, RumorCone decomposes image and text content into three levels: base, mid, and high-level abstractions, and embeds them in hyperbolic space to model their tree-like semantic structure. Second, a dynamic hyperbolic multimodal attention mechanism aligns features across modalities and levels, and a flexible fusion strategy adjusts the contribution of each modality based on alignment quality. Our experiments indicate the importance of hierarchical semantic modeling for robust and interpretable multimodal rumor detection.

Md Mahbubur Rahman, Shufeng Hao, Chongyang Shi, An Lao, Jinyan Liu (2025), The 2025 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguists (ACL)

Traditional intrusion detection systems (IDS) using classifiers face challenges such as model overfitting, incomplete feature extraction, andhighfalse positive rates, limiting their effectiveness in real-world deployments. Toaddress these challenges, this study proposes a robust multiclass machine learn ing based intrusion detection framework. The methodology integrates advanced feature selection techniques to identify critical attributes, mitigating redundancy and enhancing detection accuracy. Two distinct ML architectures are imple mented: a baseline classifier pipeline and a stacked ensemble model combining noise injection, Principal Component Analysis (PCA), and meta learning to im prove generalization and reduce false positives. Evaluated on the AWID3 dataset, the proposed ensemble architecture achieves superior performance, with an ac curacy of 98%, precision of 98%, recall of 98%, and a false positive rate of just 2%, outperforming existing state-of-the-art methods. This work demonstrates the efficacy of combining preprocessing strategies with ensemble learning to fortify network security against sophisticated Wi-Fi attacks, offering a scalable and reli able solution for IoTenvironments. Futuredirectionsincludereal-timedeployment and adversarial resilience testing to further enhance the model’s adaptability.

Rahman, M., Wang, H., Zhao, K. (2022). Network Traffic Anomaly Detection Using Hybrid Machine Learning Approaches. In Proceedings of the Workshop on Security and Privacy in Machine Learning, ACM SIGSAC, pp. 123-130.

Traditional intrusion detection systems (IDS) using classifiers face challenges such as model overfitting, incomplete feature extraction, andhighfalse positive rates, limiting their effectiveness in real-world deployments. Toaddress these challenges, this study proposes a robust multiclass machine learn ing based intrusion detection framework. The methodology integrates advanced feature selection techniques to identify critical attributes, mitigating redundancy and enhancing detection accuracy. Two distinct ML architectures are imple mented: a baseline classifier pipeline and a stacked ensemble model combining noise injection, Principal Component Analysis (PCA), and meta learning to im prove generalization and reduce false positives. Evaluated on the AWID3 dataset, the proposed ensemble architecture achieves superior performance, with an ac curacy of 98%, precision of 98%, recall of 98%, and a false positive rate of just 2%, outperforming existing state-of-the-art methods. This work demonstrates the efficacy of combining preprocessing strategies with ensemble learning to fortify network security against sophisticated Wi-Fi attacks, offering a scalable and reli able solution for IoTenvironments. Futuredirectionsincludereal-timedeployment and adversarial resilience testing to further enhance the model’s adaptability.

@article{RAHMAN2025100082,
                                    author = {Md Mahbubur Rahman and Shaharia Al Shakil and Mizanur Rahman Mustakim},
                                    title = {A survey on intrusion detection system in IoT networks},
                                    journal = {Cyber Security and Applications},
                                    volume = {3},
                                    pages = {100082},
                                    year = {2025},
                                    issn = {2772-9184},
                                    doi = {https://doi.org/10.1016/j.csa.2024.100082},
                                    url = {https://www.sciencedirect.com/science/article/pii/S2772918424000481},
                                    

Traditional rumor detection models generally rely on semantic alignment between modalities, thus suppressing modality inconsistencies as noise. However, semantic variations, such as neutral visuals paired with deceptive text, could serve as deliberate indicators of rumor. To address this problem, we propose a novel multimodal Rumor Amplification Detection And Reasoning (RADAR) framework that interprets cross-modal discrepancies as potential indicators to be amplified. RADAR contains a semantic discrepancy amplifier (SDA) module, a progressive feature purifica- tion (PFP) module, and a feature feedback loop (FFL) module. First, SDA employs an adaptive amplification technique to amplify the discrepancies between modalities and uses an attention mechanism to capture variations across semantic levels. Then PFP gradually refines the features on the different discrepancy levels to capture highly suspicious rumor content. Finally, FFL iteratively strength- ens semantic representations using discrepancy-aware purification cycles for deeper reasoning over complex rumor patterns. Extensive experiments on large-scale datasets on MR2-C, MR2-E, Weibo, and Twitter demonstrate that RADAR can detect cross-modal rumors more effectively than current techniques.

Md Mahbubur Rahman, Shufeng Hao, Chongyang Shi, An Lao, Jinyan Liu (2025), The 2025 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguists (ACL)

However, existing detection systems often struggle with accuracy, inter- pretability, and real-time efficiency, particularly in low-light conditions, complex backgrounds, and resource- constrained edge devices. To address these gaps, this study introduces three key contributions: (1) a novel homogeneous ensemble strategy that trains multiple YOLO variants with CSPDenseNet121 backbones under different frozen parameter configurations and fuses them via transfer learning, capturing hierarchical feature representations across depths, (2) the integration of visual attention blocks and Explainable AI (Grad-CAM, LIME, SHAP) to enhance feature extraction while ensuring transparency and accountability, and (3) a scalable two-stage end-to-edge deployment pipeline optimized for real-time moderation. Unlike prior heterogeneous backbone ensembles, our approach leverages a unified architecture with varied layer freezing, reducing bias and variance while maintaining computational efficiency. Robustness is further reinforced through advanced preprocessing, including normalization, class balancing, and augmentation with simulated low-light and overexposed conditions. Experimental evaluation demonstrates the effec- tiveness of the proposed framework. On the DETECTOR_AUTO_GENERATED_DATA (DAGD) dataset, our CSPDenseNet121 ensemble achieved 0.979 precision, 0.922 recall, and 92.1% accuracy, surpassing heterogeneous YOLO fusions (accuracy 89.9–91.1%) and prior baselines such as Deep CNN (91.0%). On a custom stress-test dataset, it attained 0.969 precision, 0.935 recall, and 85.0% accuracy, confirming robustness under challenging illumination and background variations. These results show that homogeneous model fusion, attention-driven mechanisms, and real-time deployment collectively advance the state of explicit content detection. By bridging high-performance AI with interpretability and scalable deployment, this research provides a comprehensive solution for modern content moderation challenges.

Wait Plz

Existing approaches are generally extracting local features (e.g., TF-IDF, CNN) or global semantics (e.g., BERT) but rarely their combination while being interpretable. We propose an explainable two stage framework for Bengali fake news detection. In the first stage, a LLaMA3 module gives formalized explanations (analysis, indicators, conclusion) of every article. These are embedded with Bangla-BERT to yield semantic vectors. In the second stage, the explanation embeddings are fused with TF-IDF and CNN based local features into a hybrid representation, enabling the model to learn both the content and its rationale. Evaluated on BanFakeNews and BFRD, our system achieves state-of-the-art performance (95.3% accuracy on BanFakeNews) while providing interpretable outputs. The system is suitable for real-time tasks in low-resource settings that are sensitive to misinformation.

Wait plz

Meanwhile, however, the problem of pornography through the Internet access in the place of work, at home and in education has considerably increased. Therefore, how effectively block or filter out pornography has been arousing a serious anxiety in related research areas. Recently, after the development of Deep Learning techniques, several algorithm have been published like as YOLO, Faster-RCNN, and Mask-RCNN, etc. In 2015, the YOLO (You Only Look Once) algorithm was evolved with a new approach, reframing object detection as a regression problem and performing in a single neural network. That made the object detection field blow up and obtained much more incredible achievements than just a decade ago. So far, combining with many of the most innovative concepts coming out of the computer vision research community, YOLO has been improved well and assessed as one of the outstanding object detection algorithms than its legacy version or others in terms of both accuracy and speed. Now, by the help of this computer vision's algorithm also can detect adult material content from images and also real-time videos. With the rapid development of Artificial Intelligence, using deep learning algorithms to detect image content, can effectively relief the workload of manual auditing, and can improve the efficiency as well. Nowadays, developers can use this algorithm in their platform to detect nudity or other's detection related issues. This thesis investigates the most advanced inventions in the field of computer vision which were integrated into YOLOv5 as well as we will use this target detection algorithms to train a detection model and deploy the model through web services, providing an auditing API services for images, and build a demo application.

Md Mahbubur Rahman,