A set of diagonal embedding algorithm for improving robustness with tamper detection in image watermarking /
Muhammad Afnan Hafidz Mohd Taslim,
A set of diagonal embedding algorithm for improving robustness with tamper detection in image watermarking / Muhammad Afnan Hafidz Mohd Taslim - xiv, 149 pages : illustrations ; 30 cm. + 1 CD-ROM.
Faculty of Computing
Thesis (Master of Science) -- Universiti Malaysia Pahang - 2026
Includes bibliographical references
Detecting suicidal intention from social media text is a challenging task because such expressions are often subtle, context-dependent, and linguistically diverse. Although Convolutional Neural Network (CNN) models have shown promise in text classification, individual CNN architectures remain limited in two important ways: they capture only specific types of spatial features according to their own architectural design, and they often show weaker generalizability across varied social media expressions. These limitations reduce their ability to represent the full range of linguistic patterns associated with suicidal intent. To address this problem, this thesis proposes a heterogeneous ensemble of CNN models for suicidal intention detection in social media text. The approach combines three distinct CNN architectures with different pooling strategies, activation functions, and attention-related mechanisms so that complementary textual features can be captured more effectively than with a single model. The study used a large, labeled Reddit dataset for model development and training, followed by evaluation against individual CNNs and other deep learning baselines. The proposed ensemble CNN achieved an accuracy of 97.56%, outperforming the individual CNN models and other compared deep learning methods. Its generalization capability was further examined using two unforeseen datasets from Reddit and Twitter, where the model maintained strong performance despite differences in platform - specific language patterns and expression styles. These findings indicate that combining heterogeneous CNN architectures can improve spatial feature extraction and strengthen classification performance for suicidal intention detection. Overall, this study contributes a CNN-based ensemble framework that provides a promising approach for identifying suicidal intention in social media text within the evaluated experimental setting.
THE0010806 (Local)
Faculty of Computing--Dissertations
Universities and colleges--Dissertations
Theses--Dissertations
A set of diagonal embedding algorithm for improving robustness with tamper detection in image watermarking / Muhammad Afnan Hafidz Mohd Taslim - xiv, 149 pages : illustrations ; 30 cm. + 1 CD-ROM.
Faculty of Computing
Thesis (Master of Science) -- Universiti Malaysia Pahang - 2026
Includes bibliographical references
Detecting suicidal intention from social media text is a challenging task because such expressions are often subtle, context-dependent, and linguistically diverse. Although Convolutional Neural Network (CNN) models have shown promise in text classification, individual CNN architectures remain limited in two important ways: they capture only specific types of spatial features according to their own architectural design, and they often show weaker generalizability across varied social media expressions. These limitations reduce their ability to represent the full range of linguistic patterns associated with suicidal intent. To address this problem, this thesis proposes a heterogeneous ensemble of CNN models for suicidal intention detection in social media text. The approach combines three distinct CNN architectures with different pooling strategies, activation functions, and attention-related mechanisms so that complementary textual features can be captured more effectively than with a single model. The study used a large, labeled Reddit dataset for model development and training, followed by evaluation against individual CNNs and other deep learning baselines. The proposed ensemble CNN achieved an accuracy of 97.56%, outperforming the individual CNN models and other compared deep learning methods. Its generalization capability was further examined using two unforeseen datasets from Reddit and Twitter, where the model maintained strong performance despite differences in platform - specific language patterns and expression styles. These findings indicate that combining heterogeneous CNN architectures can improve spatial feature extraction and strengthen classification performance for suicidal intention detection. Overall, this study contributes a CNN-based ensemble framework that provides a promising approach for identifying suicidal intention in social media text within the evaluated experimental setting.
THE0010806 (Local)
Faculty of Computing--Dissertations
Universities and colleges--Dissertations
Theses--Dissertations