Deepfake Image Detection Using Deep Learning Models

Abstract

Fake images are becoming a threat to our society. Recognizing fake images with the naked eye is a very difficult task. And in many cases it is impossible to recognize even after much effort. Due to which it is becoming easier to spread false news among people and misguide them with less effort. This problem will become even bigger in the future and will pose a threat to human safety. In this research paper I performed deep learning models one by one and from those I took the best performing models. With their help, I created a hybrid model using the best performing models. My goal is to create a hybrid model that will give high accuracy and high recall rate. And will also perform well on any custom dataset. To this end, a custom hybrid model was created and implemented leveraging innovative factorized residual blocks to achieve efficient feature extraction with reduced parameter counts. Firstly I performed with single models MobileNetV2, InceptionV3, EffieientNetV2B0, ResNet50, VGG16. From these models I took 2 models with best accuracy, precision, recall and f1 score and combined them. Then I fine-tuned for binary classification of real versus AI-generated images. This hybrid model trained and validated on the CIFAKE dataset, which contains labeled samples of authentic and synthetic images. I made a custom dataset also to check how my hybrid model on unseen dataset and it performed pretty well. Training leveraged GPU acceleration within TensorFlow/Keras frameworks to optimize computational performance. The model performance results are presented in the form of accuracy, precision, recall, f1 score, performance metrics and confusion metrics generated by the model. And I saw that Mobilentiv2 and ResNet50 are performing the same results on the CIFAKE dataset. And the rest of the models are performing almost close. So I created a hybrid model using MobilityTV2 and ResNet50 as the best performing single model. The hybrid model performs well. Robustness check is done by using a custom dataset and checking the unseen dataset. For the CIFAKE dataset my hybrid models accuracy was 0.98 and recall rate also increased to 0.98. And for the custom dataset my accuracy was 0.8875 and recall rate was 0.8875. This study underscores the feasibility of using hybrid models for fake image detection, providing a practical pathway toward scalable, real-time systems capable of mitigating the spread of AI-generated misinformation.