# Task Description
## [Background]

Generation and detection of deepfake images can provide valuable insights for various stakeholders in the fields of computer vision and cybersecurity. For example, researchers and developers participating in deepfake image competitions can use detection rate predictions to refine their models and improve performance against evolving threats. 

Additionally, predicting deepfake image generation trends can help organizers design challenging benchmarks that reflect real-world scenarios. Moreover, deepfake detection predictions can guide the development of automated systems that safeguard digital platforms from synthetic image manipulation, ensuring the integrity of online content. 

Finally, understanding trends in deepfake image generation and detection can highlight emerging challenges in digital media security and inform the development of future strategies and policies. In summary, predicting deepfake image generation and detection can play a crucial role in advancing both technological innovation and security measures in the context of deepfake image competitions.

## [Goal]

In this competition, we challenge you to predict the generation and detection rates of deepfake images. 

As part of the Red Team, you are tasked with generating deepfake images by transforming given facial images through techniques such as face swapping, expression manipulation, and mouth movement alteration. 

As part of the Blue Team, you will develop an image classification model to detect tampering and perform binary classification on the unseen test images provided by the organizers.

## [Dataset]

The dataset used in this competition is based on FaceForensics++, a collection of facial images designed for benchmarking deepfake generation and detection models.

## [Experiment Setting]

- FaceForensics++ training dataset includes all image data and corresponding labels
- Test dataset is composed of a blend of unseen public datasets and generated samples

# Submission Rules

## [Submission]

A maximum of 3 public registrations are allowed per day. Private registrations can be chosen from the results of public registrations, and if not selected, the submission will default to the highest-scoring public result. Please ensure that your team name in the leaderboard is your ID number and name.

## [Score]

• Scores are differentially assigned through a ranking competition (relative evaluation).
• Since multiple submissions are allowed, you can train the model overfitted to the test dataset. To prevent this, the test dataset is divided into an 60:40 ratio, with 60% used to calculate the public rank and the remaining 40% used to measure the private rank. The score visible on the scoreboard is the public rank, and the private rank can be viewed a day after the competition ends.

## [Report]

There is no accurate answer for modelling in machine learning, so you can try various ways to improve your model.
However, you should be able to explain how you came up with the idea, and why you used that particular method. Therefore, detailed explanation about your own analysis about the dataset, the process and your model should be in your report.

## [Rules]

Refer to 'Rules' tab for more information