Detection and Mitigation of Deepfake Attacks in Cybersecurity : Leveraging Computer Vision and Deep Learning
Zare Janakbari, Poorya (2025)
2025
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025060721436
https://urn.fi/URN:NBN:fi:amk-2025060721436
Tiivistelmä
In recent years, the production of fake content has grown exponentially, challenging cybersecurity, digital integrity, and public trust. The thesis presented here attempts to implement and evaluate a deepfake detection system that uses computer vision and deep learning techniques to classify real or manipulated facial images. In this research, three architectures are investigated (EfficientNetB4, DenseNet201, and EfficientNetV2L), each of which is a feature extractor, combined with a custom-made Dense Neural Network (DNN) that performs binary classification.
The proposed system was first trained and then tested on a balanced dataset of 140,000 images in two categories, real and fake. The system used various preprocessing and augmentation techniques to improve generalization. The system was trained on both Google Colab Pro and Pohti supercomputers to verify and ensure that the models are capable of running under different hardware conditions.
Accuracy, precision, recall, and F1-score were implemented to evaluate the system performance, which was supported by visualizations such as training curves and confusion matrices. The findings show that EfficientNetV2L + DNN has the best performance in image recognition, followed by EfficientNetB4 + DNN with remarkable accuracy. DenseNet201 + DNN faced an overfitting problem and poor performance on the test set.
The results of this study suggest that lightweight pretrained convolutional neural networks combined with dense classifiers can best detect fake content. Finally, this research proposes a comparative perspective on deep learning architectures for manipulated media analysis that can be simultaneously applied to identity verification and digital forensics.
The proposed system was first trained and then tested on a balanced dataset of 140,000 images in two categories, real and fake. The system used various preprocessing and augmentation techniques to improve generalization. The system was trained on both Google Colab Pro and Pohti supercomputers to verify and ensure that the models are capable of running under different hardware conditions.
Accuracy, precision, recall, and F1-score were implemented to evaluate the system performance, which was supported by visualizations such as training curves and confusion matrices. The findings show that EfficientNetV2L + DNN has the best performance in image recognition, followed by EfficientNetB4 + DNN with remarkable accuracy. DenseNet201 + DNN faced an overfitting problem and poor performance on the test set.
The results of this study suggest that lightweight pretrained convolutional neural networks combined with dense classifiers can best detect fake content. Finally, this research proposes a comparative perspective on deep learning architectures for manipulated media analysis that can be simultaneously applied to identity verification and digital forensics.