| Titre : | Deepfakes detection using deep learning | | Type de document : | texte manuscrit | | Auteurs : | Hadjer Koribaa, Auteur ; Leila Benarous, Directeur de thèse | | Editeur : | Laghouat : Université Amar Telidji - Département d'informatique | | Année de publication : | 2022 | | Importance : | 46 p. | | Format : | 30 cm. | | Accompagnement : | 1 disque optique numérique (CD-ROM) | | Note générale : | Option : Networks, systems and distributed applications (Réseaux,systèmes et applications réparties) | | Langues : | Anglais | | Mots-clés : | Deepfakes Classification GANs ROI Transfer learning CNNs FaceForensics++ ResNet50-V2 Deep learning | | Résumé : | Deepfakes or the hyper-realistic imitation of authentic audio-visual content, are widely spread techniques specially with the use of pre-trained generative adversarial network (GANs) that makes it easier to automatically swap a person's face with another in a video. Due to their devastating impacts on the world, distinguishing between real and deepfake videos has become a fundamental issue. Therefore, automated methods to identify these deepfake videos are required in light of recent public scandals. Much research has been devoted to developing detection methods to reduce the potential negative impact of deepfakes.
In this study, we developed an open-source platform called « Deepfake Detection », it presents a new detection technique which consists of two parts: the first part is a binary classification model that can classify videos as fake or real. The second part is another model with different input and output from the first one. It can identify which generation method among these three: FaceSwap, Face2Face and DeepFakes were used to create these deepfakes (categorical classification). We collected our samples from FaceForensics++ dataset. The preprocessing phase was necessary for this study, from frame extraction to face cropping (Region of Interest) and data augmentation. After that we split our data into train and test sets. Next, Convolutional Neural Networks (CNNs) and transfer learning approaches were employed for this work, we implemented Seven CNN-pretrained models in the first part (binary classification), with several trials and different fine-tuning parameters to determine which model is the most suitable for our situation as well as the criteria that influenced our results. The selected CNN-pretrained models are: VGG16, Inception-v3, InceptionResNet-v2, Xception, MobileNet-V2, MobileNet-V3 and ResNet50-V2. Based on the evaluation results, we reached the highest accuracy of 100% with ResNet50-V2 in both parts (the binary and categorical classification). Lastly, we developed our web application, for users to interact with our models. | | note de thèses : | Mémoire de master en informatique |
Deepfakes detection using deep learning [texte manuscrit] / Hadjer Koribaa, Auteur ; Leila Benarous, Directeur de thèse . - Laghouat : Université Amar Telidji - Département d'informatique, 2022 . - 46 p. ; 30 cm. + 1 disque optique numérique (CD-ROM). Option : Networks, systems and distributed applications (Réseaux,systèmes et applications réparties) Langues : Anglais | Mots-clés : | Deepfakes Classification GANs ROI Transfer learning CNNs FaceForensics++ ResNet50-V2 Deep learning | | Résumé : | Deepfakes or the hyper-realistic imitation of authentic audio-visual content, are widely spread techniques specially with the use of pre-trained generative adversarial network (GANs) that makes it easier to automatically swap a person's face with another in a video. Due to their devastating impacts on the world, distinguishing between real and deepfake videos has become a fundamental issue. Therefore, automated methods to identify these deepfake videos are required in light of recent public scandals. Much research has been devoted to developing detection methods to reduce the potential negative impact of deepfakes.
In this study, we developed an open-source platform called « Deepfake Detection », it presents a new detection technique which consists of two parts: the first part is a binary classification model that can classify videos as fake or real. The second part is another model with different input and output from the first one. It can identify which generation method among these three: FaceSwap, Face2Face and DeepFakes were used to create these deepfakes (categorical classification). We collected our samples from FaceForensics++ dataset. The preprocessing phase was necessary for this study, from frame extraction to face cropping (Region of Interest) and data augmentation. After that we split our data into train and test sets. Next, Convolutional Neural Networks (CNNs) and transfer learning approaches were employed for this work, we implemented Seven CNN-pretrained models in the first part (binary classification), with several trials and different fine-tuning parameters to determine which model is the most suitable for our situation as well as the criteria that influenced our results. The selected CNN-pretrained models are: VGG16, Inception-v3, InceptionResNet-v2, Xception, MobileNet-V2, MobileNet-V3 and ResNet50-V2. Based on the evaluation results, we reached the highest accuracy of 100% with ResNet50-V2 in both parts (the binary and categorical classification). Lastly, we developed our web application, for users to interact with our models. | | note de thèses : | Mémoire de master en informatique |
|
Ajouter le résultat dans votre panier Faire une suggestion Affiner la rechercheDeepfakes detection using deep learning / Hadjer Koribaa
