|
|
Use of Diffusion Models in Deepfakes
Trúchly, Dominik ; Malinka, Kamil (oponent) ; Lapšanský, Tomáš (vedoucí práce)
A deepfake is a type of synthetic media created through sophisticated machine learning algorithms, particularly deep neural networks. As an example Generative adversarial neural networks (GANs), that are capable of generating images that are almost impossible for ordinary individuals to differentiate from genuine reality. Consequently, deepfake detection algorithms have been developed to address this growing concern. Leveraging advanced machine learning techniques, these algorithms analyze various features within images and videos to identify inconsistencies or anomalies indicative of manipulation. This thesis investigates the application of diffusion models, commonly utilized in digital image processing to enhance image quality by reducing noise and blurring, in bolstering the realism of deepfakes. By using these models, we test their effect on detecting deepfakes images using deepfake detectors.
|
|
|
Diffusion Models and their Impact on Cybersecurity
Dvorščák, Patrik ; Homoliak, Ivan (oponent) ; Lapšanský, Tomáš (vedoucí práce)
This thesis explores the performance of diffusion models (DMs) and generative adversarial networks (GANs) in creating AI-generated visual content across multiple applications, including face synthesis, text-to-image generation, artistic rendering, image-to-image translation, video synthesis, and super-resolution. Through comparative experiments, this research evaluates the models' ability to generate detailed, realistic, and artistically compelling visuals from textual and image prompts. The results reveal that DMs excel in producing highly detailed images that closely follow text prompts, particularly effective in face synthesis and text-to-image tasks. In contrast, GANs are more adept at rendering realistic environmental scenes, suitable for applications requiring immersive visuals. Both model types are competent in artistic rendering, though they differ in style adaptation and creativity. The thesis concludes with future research directions aimed at enhancing model efficacy and integrating these technologies more effectively into practical applications.
|
|
|
Creating Novel Deepfake Speech Dataset
Sztolarik, Maroš ; Homoliak, Ivan (oponent) ; Firc, Anton (vedoucí práce)
In the recent years, deepfake technology has advanced to a point where it can convincingly mimic human speech, posing significant challenges in distinguishing between real and synthetic voices. In this thesis, we introduce a novel dataset comprising speech deepfakes generated using diffusion models. This dataset, created with two sophisticated text-to-speech tools, DiffSpeech and ProDiff, aims to provide insight into the threat that these new tools pose. Two more datasets are created with more mature tools, Glow-TTS and Tacotron2, to provide a point of comparison. Then all the generated samples are analyzed through two deepfake detectors in order to provide a direct comparison into how much of a threat each tool is to these detectors. The results show that even though the tools utilizing the diffusion models are threatening, the use of diffusion models did not provide these tools any meaningful advantage in evading the detection.
|
|
|
Generative Models for 3D Shape Completion
Zdravecký, Peter ; Španěl, Michal (oponent) ; Kubík, Tibor (vedoucí práce)
In many real-world scenarios, scanned 3D models contain missing parts due to occlusion, scanning errors, or the incomplete nature of the data itself. The goal of this work is to create an automated process for 3D shape completion using a supervised deep learning-based method. The proposed solution is based on the prior work of DiffComplete, which uses a diffusion-based model operating over distance field representation and handles the task as a generative problem. The results showed a high capability of this model with an 81.6 IoU metric on the custom-prepared test set of furniture objects. The model also demonstrates strong generalization capabilities on shapes that are out of the training distribution (average 70.9 IoU metric). Apart from more detailed data-centric experiments, this work further extends current state-of-the-art in two ways. Firstly, it addresses the most crucial shortcoming, expensive computation, by processing the input in a low-resolution domain. Secondly, it utilizes user input (Region of Interest), which gives the user more control over generation in ambiguous scenarios.
|
|
|
Model Compression of Denoising Diffusion Probabilistic Models for Image Generation
Dobiš, Lukáš ; Kišš, Martin (oponent) ; Hradiš, Michal (vedoucí práce)
This thesis is dedicated to optimizating computation in generative diffusion models by evaluating conventional model compression aproaches on Denoising Diffusion Probabilistic Model (DDPM). Model compression was done on parameters of pretrained baseline DDPM neural network by several quantization and pruning methods. These methods were evaluated on three image dataset benchmarks. Results show that quantization and pruning are viable compression methods for downsizing diffusion networks, because they showed little decrease in quality of generated imagery. These results confirm that implemented compression approaches are vital for deployment of diffusion models on resource constrained Edge devices or to offset their compute costs.
|
host ::
RSS.