Department of Electrical and Computer Engineering, North South University, Dhaka 1229, Bangladesh.
School of Computing and IT, Manipal University Jaipur, Jaipur, Rajasthan, India.
Comput Intell Neurosci. 2021 Dec 16;2021:3111676. doi: 10.1155/2021/3111676. eCollection 2021.
Generation Z is a data-driven generation. Everyone has the entirety of humanity's knowledge in their hands. The technological possibilities are endless. However, we use and misuse this blessing to face swap using deepfake. Deepfake is an emerging subdomain of artificial intelligence technology in which one person's face is overlaid over another person's face, which is very prominent across social media. Machine learning is the main element of deepfakes, and it has allowed deepfake images and videos to be generated considerably faster and at a lower cost. Despite the negative connotations associated with the phrase "deepfakes," the technology is being more widely employed commercially and individually. Although it is relatively new, the latest technological advances make it more and more challenging to detect deepfakes and synthesized images from real ones. An increasing sense of unease has developed around the emergence of deepfake technologies. Our main objective is to detect deepfake images from real ones accurately. In this research, we implemented several methods to detect deepfake images and make a comparative analysis. Our model was trained by datasets from Kaggle, which had 70,000 images from the Flickr dataset and 70,000 images produced by styleGAN. For this comparative study of the use of convolutional neural networks (CNN) to identify genuine and deepfake pictures, we trained eight different CNN models. Three of these models were trained using the DenseNet architecture (DenseNet121, DenseNet169, and DenseNet201); two were trained using the VGGNet architecture (VGG16, VGG19); one was with the ResNet50 architecture, one with the VGGFace, and one with a bespoke CNN architecture. We have also implemented a custom model that incorporates methods like dropout and padding that aid in determining whether or not the other models reflect their objectives. The results were categorized by five evaluation metrics: accuracy, precision, recall, 1-score, and area under the ROC (receiver operating characteristic) curve. Amongst all the models, VGGFace performed the best, with 99% accuracy. Besides, we obtained 97% from the ResNet50, 96% from the DenseNet201, 95% from the DenseNet169, 94% from the VGG19, 92% from the VGG16, 97% from the DenseNet121 model, and 90% from the custom model.
Z 世代是一个数据驱动的世代。每个人都掌握着整个人类的知识。技术的可能性是无穷无尽的。然而，我们却在误用这个恩赐，利用深度伪造技术进行换脸。深度伪造是人工智能技术的一个新兴子领域，其中一个人的脸被叠加在另一个人的脸上，这在社交媒体上非常突出。机器学习是深度伪造的主要元素，它使得深度伪造图像和视频的生成速度更快，成本更低。尽管“深度伪造”这个词带有负面含义，但该技术在商业和个人领域的应用越来越广泛。尽管它相对较新，但最新的技术进步使得从真实图像中检测深度伪造和合成图像变得越来越具有挑战性。随着深度伪造技术的出现，人们越来越感到不安。我们的主要目标是准确地从真实图像中检测深度伪造图像。在这项研究中，我们实施了几种方法来检测深度伪造图像并进行比较分析。我们的模型是通过 Kaggle 数据集进行训练的，这些数据集来自 Flickr 数据集的 70000 张图像和由 styleGAN 生成的 70000 张图像。为了对卷积神经网络（CNN）用于识别真实和深度伪造图片的方法进行比较研究，我们训练了八个不同的 CNN 模型。其中三个模型是使用 DenseNet 架构（DenseNet121、DenseNet169 和 DenseNet201）训练的；两个是使用 VGGNet 架构（VGG16、VGG19）训练的；一个是使用 ResNet50 架构的，一个是使用 VGGFace 的，一个是使用定制的 CNN 架构的。我们还实现了一个自定义模型，该模型结合了 dropout 和 padding 等方法，以确定其他模型是否反映了它们的目标。结果按五个评估指标进行分类：准确率、精度、召回率、1 分和 ROC 曲线下的面积。在所有模型中，VGGFace 的表现最好，准确率达到 99%。此外，我们从 ResNet50 获得了 97%，从 DenseNet201 获得了 96%，从 DenseNet169 获得了 95%，从 VGG19 获得了 94%，从 VGG16 获得了 92%，从 DenseNet121 模型获得了 97%，从自定义模型获得了 90%。
