Related Experiment Video
Updated: Jun 21, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
515
Detection of fake face images using lightweight convolutional neural networks with stacking ensemble learning method
Emre Şafak1,2, Necaattin Barışçı2
1R&D Technology and Innovation Department, HAVELSAN, Ankara, Türkiye.
Peerj. Computer Science
|July 10, 2024
Summary
This study developed a machine learning model to detect fake face images. An ensemble learning method achieved 96.44% accuracy, improving upon individual convolutional neural networks for identifying manipulated media.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Digital Forensics
Background:
- Digital manipulation of images and videos, particularly fake faces, is prevalent and poses risks of misinformation.
- Machine learning, especially generative adversarial networks (GANs), increasingly complicates the detection of synthetic media.
- Face manipulations encompass entire face synthesis, deepfakes, attribute, and expression manipulation.
Purpose of the Study:
- To investigate the effectiveness of lightweight convolutional neural networks (CNNs) in detecting fake face images.
- To compare the performance of various CNN architectures, including MobileNet, MobileNetV2, EfficientNetB0, and NASNetMobile.
- To enhance detection accuracy through model revision and ensemble learning techniques.
Main Methods:
- Utilized a dataset of 70,000 real (FFHQ) and 70,000 fake (StyleGAN2-generated) face images for training and testing (80%/20% split).
- Employed transfer learning by pre-training MobileNet, MobileNetV2, EfficientNetB0, and NASNetMobile on ImageNet.
- Implemented model refinement by adding dense and dropout layers to EfficientNetB0, followed by stacking ensemble learning with MobileNet and MobileNetV2.
Main Results:
- Initial training showed EfficientNetB0 achieving the highest accuracy at 93.64%.
- Post-revision, EfficientNetB0 reached an accuracy of 95.48%.
- The final stacking ensemble model, combining EfficientNetB0, MobileNet, and MobileNetV2, achieved a peak accuracy of 96.44%.
Conclusions:
- Lightweight CNNs, particularly when enhanced and combined through ensemble learning, are effective for detecting synthetic face images.
- The developed ensemble model demonstrates superior performance in distinguishing real from fake faces compared to individual models.
- This research contributes to combating misinformation by providing a more accurate method for identifying manipulated visual content.
