## Background
- Techniques for convincing manipulations of digital facial images
exist for decades via the use of visual effects 
- Recent technologies in deep learning, especially “DeepFakes,”
have led to a dramatic increase in the realism of fake content
- While there are entertaining applications of DeepFakes, potential
weaponization of such techniques has raised great concerns
- DeepFake detection has become a pressing issue and a hot spot of
research

## Task: 
A binary classification problem
- Input: RGB facial images of size 299 x 299 x 3
- Output: A binary label to indicate whether an image is fake or not

## Training set: 
12,000 images, within which 8,000 images contain
fake faces, and the rest images are real
- Download link: https://drive.google.com/drive/folders/1n18AqgUF_KqKbcxuLXrTcLq-pyfhSJD9?usp=sharing

## Test set: 
3,000 images that will not be released to the students
- Students have THREE chances to get access to the test set
- The best result among the three will be used for ranking

- Tip: It is a good practice to divide the available set further into
training, validation, and test set

## Group project: 
A group of at most two students is allowed

## What to hand in: 
a single ipython notebook (.ipynb) as the project report with source code files included
- Source files must contain the training code for reproduction
- An extra PDF file as project report is not recommended

## When to hand in: 
Dec. 6, 2021, 11:59 pm

## Where to hand in: 
- Submit the clean and runable test code to the TAs
- Submit the .ipynb file via Canvas

## GPUs: 
Some GPU hours are available on Microsoft Azure

## Grading (Totally 30 points)
- 30.0% - Technical correctness (whether the
methodologies/algorithms are correctly used)
- 30.0% - Experiment and analysis
- 20.0% - Quality of the written report (organized, complete
descriptions, etc.)
- 10.0% - Quality of project presentation (tentatively held in Wed,
Nov. 24)
- 10.0% - Reserved for Top-5 teams based on the test set
performance