Browsing by Subject "Deep Learning"

Now showing 1 - 4 of 4
  • No Thumbnail Available
    Item
    Open Access
    Deep learning models for malware traffic detection and classification: a focus on pruning and hyperband for resource optimization
    (2026) Sheetekela, Sanette Penelao; Chavula, Josiah
    Network traffic encryption is increasingly being adopted as the norm for communication technologies because of its capacity to protect privacy. Conversely, concealing data from unintended recipients presents an additional issue in the communication system, as encrypted data makes it challenging to detect malware using traditional methods such as deep packet inspection. Deep learning (DL) has emerged as a powerful technique for detecting malware in encrypted traffic flows. However, their high processing needs make it hard to use them in places with limited resources, such as Internet of Things (IoT) and edge devices. This study investigated the effectiveness of post-training weight and neuron pruning on three DL models—Multilayer Perceptron (MLP), One-dimensional convolutional Neural Networks (1D-CNN), and Long Short-term Memory Networks (LSTM)—to optimize resource usage while maintaining detection accuracy. In addition, the study integrated Hyperband, a resource-efficient hyperparameter tuning method that dynamically allocates computational resources based on intermediate performance evaluations to enhance pruned models. The CTU-13 dataset was used to assess how pruning rates influence detection performance and resource efficiency, with an emphasis on the trade-offs introduced by weight pruning (WP) and neuron pruning (NP). Our findings revealed that moderate pruning of 20% to 40% maintained or improved model accuracy while decreasing the model size by up to 40%, inference time by 30%, and memory usage by 10%. Among the models, 1D-CNN consistently provided the optimal balance between performance and computational efficiency. Regarding pruning techniques, structured neuron pruning (NP) has proven to be the most effective method, achieving significant reductions in model size and memory usage without compromising on detection performance. In contrast, unstructured weight pruning (WP) primarily reduced the number of active weights in the model by setting them to zero, which helped lower memory usage during processing; however, no compression was applied to the actual file size. Overall, the findings showed that adaptive tuning, when combined with pruning, offers an effective approach for building high-performing lightweight models that can be implemented in resource-constrained environments for encrypted malware traffic flow classification.
  • No Thumbnail Available
    Item
    Open Access
    Enhancing detection of cervical cancer through deep learning: a comparative study of histological image-based algorithms
    (2025) Tjale, Palesa; Sinkala, Musalula
    Cervical cancer is a significant contributor to cancer-related deaths among women worldwide, especially in low- and middle-income countries (LMICs) where access to screening services is limited. Early detection plays a vital role in improving patient outcomes. However, traditional diagnostic techniques, including Pap smears and histological assessments, are often affected by variability, subjectivity, and limited sensitivity. Advances in artificial intelligence (AI), particularly deep learning (DL) and visual prompting methods, offer new possibilities for enhancing the accuracy, efficiency, and interpretability of cervical cancer detection from histology images. In this thesis, I investigate the application of DL models—ResNet50, SqueezeNet, EfficientNet, and a Visual Prompting Model—for classifying cervical cells using histopathological images. I conduct a comparative analysis to evaluate these models based on accuracy, sensitivity, specificity, and interpretability. To enhance model explainability, I employ Grad-CAM to visualize model decisions, offering insights into the diagnostic relevance of highlighted features. My results indicate that the Visual Prompting Model outperforms conventional DL models, achieving the highest accuracy (98%) and F1-score (0.99) while also demonstrating superior localization of diagnostically significant regions. EfficientNet follows closely with an accuracy of 97% and an F1-score of 0.97, while SqueezeNet achieves 95% accuracy and an F1-score of 0.95. In contrast, ResNet50 shows lower performance, with an accuracy of 91% and an F1-score of 0.91, indicating limitations in feature extraction and localization. A key finding of my study is that integrating visual prompting significantly enhances model explainability, addressing a critical challenge in AI-driven medical imaging. By directing attention to clinically relevant areas within histological images, visual prompting reduces misclassification rates, potentially aiding pathologists in making more informed diagnostic decisions. Additionally, the computational efficiency and ease of training of Visual Prompting Models suggest their feasibility for deployment in resource-constrained settings where expert pathology review is limited. Overall, my findings underscore the transformative potential of AI, particularly visual prompting, in improving cancer detection. These AI-assisted diagnostic tools promise not only to enhance accuracy but also to improve interpretability, making them highly relevant for clinical integration. I suggest that future research should focus on validating these AI models across diverse clinical settings, optimizing computational efficiency, and exploring hybrid AI approaches that incorporate molecular and genomic data for a more comprehensive approach to cervical cancer diagnostics.
  • No Thumbnail Available
    Item
    Open Access
    Evaluating convolutional neural networks and transformer architectures for image-based prediction of protein localization in eukaryotic cells
    (2025) Msipa, Sibongiseni Letticia; Sinkala, Musalula
    Background: Accurate prediction of protein subcellular localization is critical for understanding protein function and guiding experimental research. Recent advances in deep learning have enabled high-throughput image-based methods to tackle this problem by leveraging large-scale immunofluorescence microscopy datasets. The aim of this study is to comparatively evaluate convolutional neural network (CNN) architectures and Transformer- based models for the multi-label classification of protein subcellular localization in eukaryotic cells, using large-scale immunofluorescence image datasets. Methods: In this study, we comparatively evaluated convolutional neural network (CNN) architectures (DenseNet121, Xception, and InceptionV3) and transformer-based models (Vision Transformer and Swin Transformer) for multi-label classification of protein localization in eukaryotic cells. Using 12,565 immunofluorescence images from the Human Protein Atlas—representing 15 subcellular compartments—we performed transfer learning by replacing the final layers of pretrained ImageNet models to accommodate multi-label output. All models were trained with iterative stratification to handle class imbalance and evaluated on held-out test images. Results and discussion: Our findings indicate that CNN-based models, particularly DenseNet121 and Xception, achieve the highest overall accuracy and F1-scores, successfully recognizing both abundant and underrepresented classes. In contrast, transformers demonstrated variable performance. While the Swin Transformer surpassed the Vision Transformer, neither consistently matched CNN performance—likely reflecting the data requirements and hyperparameter sensitivity of transformer architectures. Visualization techniques (Grad-CAM in CNNs and attention maps in transformers) confirmed that well- performing models localize salient features to biologically relevant regions, suggesting they learn meaningful morphological cues Conclusion: These results underscore CNNs' suitability for subcellular localization analysis with moderate-scale datasets, while transformers may require more extensive tuning or larger training sets to reach comparable accuracy. Our findings suggest that CNNs, especially DenseNet121 and Xception, exhibit superior performance over transformer models in predicting protein localization. CNN-based models demonstrate higher accuracy and interpretability, positioning them as preferred choices for advancing functional proteomics and computational drug discovery.
  • No Thumbnail Available
    Item
    Open Access
    Wildfire path spread prediction system using machine learning: use of ANN, convolutional autoencoder, and ConvLSTM ML model
    (2025) Makhaba, Limpho Mapulane; Winberg, Simon
    Over the years, the frequency and intensity of wildfires have increased due to climate change, old fire management practices, and other environmental factors [1]. These wildfires pose a threat to ecosystems, infrastructure, property, and lives, particularly in highly susceptible Mediterranean regions, with suitable climatic conditions characterized by hot, dry summers and cool, humid winters [2]. Although fire is essential to maintaining such ecosystems' health, it can destroy long-term environmental sustainability and cause devastating destruction [1, 3]. Wildfire spread prediction from the ignition point to the surrounding area is a complex phenomenon affected by multiple variables such as weather, topography, and land-cover characteristics. Traditional systems used in wildfire spread prediction are often limited in accounting for the dynamic and complex factors influencing fire behaviour. To address this challenge, this research uses a multidisciplinary approach that combines Machine Learning techniques, remote sensing data, and wildfire science to develop an advanced wildfire prediction system. This research performs a comprehensive investigation into the use of supervised classification Machine Learning(ML) models, specifically, Artificial Neural Networks(ANN), Convolutional Autoencoder, and Convolution Long Short-Term Memory(convLSTM), in predicting wildfire spread. The three ML models developed were based on domain knowledge of fire behaviour and utilised Google Earth Engine(GEE) weather data and Sentinel Hub burned scars satellite data to train a logistic method for simulating fire burn maps. The ML models were then trained, tested, and validated using this data. An analysis of the models' capabilities in predicting the propagation of wildfire spread, as well as the driving factors that affect wildfire behaviour, was performed. The following evaluation matrices: loss, precision, recall, Area Under the Curve-Precision Recall(AUC_PR) and F1 score are used to assess the performance of the models in predicting the spatial and temporal properties of wildfire spread. The results indicate that the convLSTM can predict both temporal and spatial properties of fire, while the Convolutional Autoencoder can predict only the spatial properties of fire with minimal loss. The ANN produced the least satisfying results in predicting fire's spatial properties.