Recent years have seen notable progress in generative modeling, leading to the emergence of the Triple-BigGAN model as an extension of the pioneering BigGAN model. This thesis scrutinizes the Triple-BigGAN model, investigating its role in augmenting image quality and its integration with a co-trained classifier. Through comprehensive experimentation and analysis, this research probes into the complexities encountered during experimentation and the insights gleaned from employing varied computational environments such as Google Colab, Kaggle Notebooks, and Google Vertex AI. Utilizing metrics like Fréchet Inception Distance (FID), Inception Score (IS), categorical cross-entropy loss, and accuracy, the dissertation evaluates the algorithm's efficacy in both image generation and classification tasks. It delineates the delicate balance among the generator, discriminator, and classifier elements within the model. This exploration of the Triple-BigGAN algorithm contributes to a deeper comprehension of advanced generative models, elucidating their potentials and challenges while laying the groundwork for further advancements at the nexus of generative and discriminative AI techniques.
3 February 2025 in Iran Journal of Computer Science
This study delves into unsupervised handwritten signature verification, aiming to develop effective techniques without relying on labeled data. To address the challenge of limited training data, the researchers creatively employ an extensive artificially generated dataset containing signatures from over 3,000 unique individuals. The primary objective is to explore the potential of Extreme Learning Machines (ELMs) in offline signature verification, leveraging innovative unsupervised methodologies. The methodology involves several steps, including the generation of high-quality signature images, extraction of local features using a sliding window technique, and utilization of deep learning principles to train a comprehensive ELM model. The final image prediction is obtained by aggregating predictions made on local features, followed by thorough performance evaluation and assessment of practical applicability. Through meticulous methodology and rigorous evaluation, this study demonstrates a novel approach to unsupervised handwritten signature verification. By showcasing the efficacy of utilizing artificial datasets and ELMs, the research provides valuable insights for advancing signature verification systems, with potential implications for enhancing security measures across various domains.
The study addresses the pressing need for improved malware detection in cybersecurity, leveraging a novel approach that combines deep learning with feature selection techniques. By analyzing network traffic patterns, the research aims to distinguish between benign and malicious activities, crucial for protecting networks and devices from cyber threats. Through training and evaluating Dense and LSTM neural network models on extensive malware datasets, the study demonstrates a significant enhancement in detection accuracy, surpassing previous methodologies by 0.98%. This innovative method not only provides valuable insights for developers but also contributes to advancing the capabilities of malware detection systems, ultimately bolstering cybersecurity defenses against evolving threats.