Detection of Image Level Forgery with Various Constraints Using DFDC Full and Sample Datasets.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Advanced machine learning techniques like autoencoders and generative adversarial networks create realistic deepfake images.
• Deepfakes are increasingly used unethically, posing challenges for distinguishing real from fake imagery.
• Existing deepfake detection methods often lack flexibility in handling diverse image resolutions and aspect ratios.

Purpose of the Study:

• To propose and evaluate a lightweight convolutional neural network (CNN) for accurate deepfake image detection.
• To assess the model's performance on both sample and full Deep Fake Detection Challenge (DFDC) datasets.
• To explore the model's adaptability to various image resolutions and aspect ratios, a novel approach for DFDC datasets.

Main Methods:

• Development of a lightweight convolutional neural network (CNN) architecture.
• Training and testing the proposed CNN model using the Deep Fake Detection Challenge (DFDC) sample and full datasets.
• Comparative analysis against state-of-the-art pre-trained models (VGG-19, Xception, Inception-ResNet-v2) and experimentation with 1:1 and 9:16 aspect ratios.

Main Results:

• The proposed lightweight CNN model achieves effective deepfake detection capabilities.
• The model demonstrates flexibility in handling various image resolutions and aspect ratios, outperforming constraints of previous studies.
• Performance evaluation on full DFDC datasets provides comprehensive insights beyond sample-based analyses.

Conclusions:

• The developed lightweight CNN offers a robust and adaptable solution for deepfake image detection.
• The model's ability to process diverse resolutions and aspect ratios expands its applicability in real-world scenarios.
• This research contributes a novel deepfake detection methodology, particularly for DFDC datasets, paving the way for future advancements.