BATTLING DEEP FAKE DETECTION USING DEEP LEARNING

 

                                                                     ABSTRACT

• Deepfake technology, which can create highly realistic yet fake videos, has sparked concern due to its potential to spread misinformation, violate privacy, and facilitate fraud. While this technology holds promise for entertainment and education, it also presents real risks. Our project aims to address these issues by developing an advanced, user-friendly deepfake detection tool.  

• Our project addresses the growing risk of deepfakes by creating a user-friendly tool to help detect manipulated media. We trained AI models on a dataset of real and altered videos, refining the model's accuracy using techniques like transfer learning with VGG16, ResNet50, and EfficientNet. To make detection accessible, we built an intuitive web interface where users can upload images or videos to check for deepfakes in real time. This tool offers a practical way to spot fake media and supports efforts to keep digital spaces trustworthy.

                                                        ARCHITECTURE DIAGRAM
                                                                   

                                             

                                                                PROPOSED SYSTEM      

To enhance our deepfake detection, we used transfer learning models like VGG16, ResNet50, and EfficientNet, building on their powerful, pre-learned features to accurately identify real versus fake faces. We also created a hybrid model that combines EfficientNet and ResNet50, taking advantage of each model’s strengths to further improve detection accuracy.

For training, we used binary cross-entropy, a loss function well-suited to distinguishing real from fake content, and optimized the model with Adam to speed up learning. Throughout training, we monitored accuracy and loss to ensure the model’s performance stayed on track.

                                                                      

                                                                     RESULTS

                    

To train our deepfake detection model, we divide the data into smaller batches to optimize memory usage. We iteratively train the model over multiple epochs, adjusting the learning rate to ensure efficient convergence. To prevent overfitting, we monitor the model's performance on a validation set and stop training early if performance starts to decline.

Evaluating the Model's Performance

Once training is complete, we evaluate the model's performance on a separate test dataset. Key metrics we use include:

Accuracy, Confusion Matrix, precision, recall, and F1-score for each class.

We compare the performance of different model architectures (VGG16, ResNet50, EfficientNet, and a hybrid) to select the best-performing one.

Integrating the Model into a Web Application

To make our deepfake detector accessible to users, we've built a web application. The backend, powered by Flask, handles user uploads, processes them using the selected model, and sends the results back to the frontend.The user-friendly frontend, built with HTML, CSS, and JavaScript, allows users to upload videos or images. Once the model has analyzed the input, the results are displayed on the webpage, highlighting real and fake frames.