Fake image detection using convolutional neural networks and hybrid machine learning techniques.
(2025) DABN01 20251Department of Economics
Department of Statistics
- Abstract
- The rise of manipulated visual content on digital platforms (whether through artificial
intelligence, photo editing software, or simple filters) has led to growing concerns
around misinformation. This thesis investigates the use of deep learning models,
specifically Convolutional Neural Networks (CNNs), to detect fake images. Several
architectures derived from pretrained models (in which we added some layers),
including EfficientNet-B0, B1, B2, and MobileNet-V3, were implemented and fine-tuned
for binary classification. The models were trained and evaluated on a balanced dataset
of real and fake images. Among them, EfficientNet-B2 achieved the best performance,
reaching a test accuracy of 92%, demonstrating strong generalization.... (More) - The rise of manipulated visual content on digital platforms (whether through artificial
intelligence, photo editing software, or simple filters) has led to growing concerns
around misinformation. This thesis investigates the use of deep learning models,
specifically Convolutional Neural Networks (CNNs), to detect fake images. Several
architectures derived from pretrained models (in which we added some layers),
including EfficientNet-B0, B1, B2, and MobileNet-V3, were implemented and fine-tuned
for binary classification. The models were trained and evaluated on a balanced dataset
of real and fake images. Among them, EfficientNet-B2 achieved the best performance,
reaching a test accuracy of 92%, demonstrating strong generalization. In addition,
traditional machine learning algorithms such as SVM, Random Forest, and Logistic
Regression were also applied. On one hand, features extracted directly from the CNN
model were used with these algorithms, achieving competitive results. On the other
hand, a set of handcrafted features including color, texture, noise, edges, compression,
and landmarks was extracted from the raw images to evaluate their generalization
capabilities and to assess whether they could enhance performance compared to CNN
derived features. These findings confirm the effectiveness of CNNs in detecting
manipulated visual content and highlight the importance of model selection and dataset
characteristics in building reliable detection systems. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9193558
- author
- Bernal Aparicio, Alba LU and Sainero Valle, David
- supervisor
- organization
- course
- DABN01 20251
- year
- 2025
- type
- H1 - Master's Degree (One Year)
- subject
- keywords
- fake images, deep learning, CNN, image classification.
- language
- English
- id
- 9193558
- date added to LUP
- 2025-09-12 09:03:37
- date last changed
- 2025-09-12 09:03:37
@misc{9193558, abstract = {{The rise of manipulated visual content on digital platforms (whether through artificial intelligence, photo editing software, or simple filters) has led to growing concerns around misinformation. This thesis investigates the use of deep learning models, specifically Convolutional Neural Networks (CNNs), to detect fake images. Several architectures derived from pretrained models (in which we added some layers), including EfficientNet-B0, B1, B2, and MobileNet-V3, were implemented and fine-tuned for binary classification. The models were trained and evaluated on a balanced dataset of real and fake images. Among them, EfficientNet-B2 achieved the best performance, reaching a test accuracy of 92%, demonstrating strong generalization. In addition, traditional machine learning algorithms such as SVM, Random Forest, and Logistic Regression were also applied. On one hand, features extracted directly from the CNN model were used with these algorithms, achieving competitive results. On the other hand, a set of handcrafted features including color, texture, noise, edges, compression, and landmarks was extracted from the raw images to evaluate their generalization capabilities and to assess whether they could enhance performance compared to CNN derived features. These findings confirm the effectiveness of CNNs in detecting manipulated visual content and highlight the importance of model selection and dataset characteristics in building reliable detection systems.}}, author = {{Bernal Aparicio, Alba and Sainero Valle, David}}, language = {{eng}}, note = {{Student Paper}}, title = {{Fake image detection using convolutional neural networks and hybrid machine learning techniques.}}, year = {{2025}}, }