|
|
 |
| |
|
|
|
Deepfake Image Forensics: Towards Efficient and Reliable Detection |
|
|
|
PP: 341-349 |
|
|
|
doi:10.18576/isl/130212
|
|
|
|
Author(s) |
|
|
|
Manar M. Hafez,
Mamdouh Mohamed,
Youssef Hesham,
Mahmoud M. Gomaa,
Ghidan Sadek,
Alyaa Amer,
|
|
|
|
Abstract |
|
|
| In an era marked by the exponential growth and sophistication of deepfakes, heightened concerns among the public about the increasing role of these manipulative digital creations in disseminating disinformation have become evident. The surge in AI-driven research focusing on deepfakes, spanning creation methods, detection techniques, and datasets, underscores the urgency to address the challenges posed by these deceptive technologies. While deep learning methods have proven successful in identifying and preventing deepfakes, the rapid evolution of deep-fake technology continues to present critical challenges, especially in the realm of image forgery detection. To address these challenges head-on, our paper introduces an advanced deepfake image forensics approach leveraging Convolutional Neural Networks (CNN), incorporating ResNet50, DenseNet121, and InceptionV3. These three prominent deep learning models are strategically combined in an innovative ensemble stacking methodology, surpassing traditional approaches to significantly enhance detection performance. High-level features are meticulously extracted from two distinct base models, DenseNet121 and InceptionV3, which are then synergistically combined and input into a Logistic Regression (LR) meta-model, serving as a robust classifier for final predictions. To ensure the reliability of our proposed approach, we conducted extensive training and fine-tuning using a large dataset comprising 70,000 real human faces from the Flickr dataset. This dataset was enriched with an additional 70,000 manipulated faces generated through the style-GAN technique. Through comprehensive evaluation, our approach demonstrated an exceptional accuracy of 98.7% in detecting image forgeries. Additionally, the Error Level Analysis (ELA) technique was applied to the dataset and employed across the three deep learning models. This advanced deepfake image forensics system stands as a significant contribution to ongoing efforts in combating deepfake risks. By empowering users to authenticate digital images and detect potential manipulations, it plays a pivotal role in safeguarding against misinformation and fraudulent activities.
|
|
|
|
|
|
|
|
