Determining Crowd’s Ability to Distinguish Deepfakes in Images and Videos

This project was an effort to determine whether a crowd has the ability to distinguish Deepfake images and videos from real images and videos. This was submitted for the final project as part of the CS395T Human Computation and Crowdsourcing course at UT Austin.

Background

In the modern era where Machine Learning (ML) and Artificial Intelligence (AI) are taking over a lot of fields, we tend to forget the large amounts of data that is involved in training these algorithms. To top it off, a huge portion of the algorithms require supervised learning where the data needs to be labeled with the correct output the algorithm needs to learn. With millions or billions of data samples to solve a single problem, data labeling has become a tedious task.

Over the years, a new way to fulfill data labeling needs took shape. Called "Crowdsourcing", this new method involved outsourcing the data labeling work to other people who are willing to do such repetitive work for some amount of money per data sample labeled. However, considering humans have a tendency to be erroneous and sometimes even malicious, crowdsourcing has become a field of study in itself. This includes understanding human behavior, learning what tasks humans are good at, learning how to maximize the output quality and quantity, how to curb mal-intent, and a lot of other areas.

A lot of platforms became quite large since the concept of crowdsourcing became popular, and for some people, these platforms became a main source of income. These platforms allow requesters to create Human Intelligence Tasks (HITs) that the workers could then work on and earn money. The largest single platform for crowdsourcing is Amazon Mechanical Turk (MTurk), which was launched in 2005. More recently, the popularity further increased when Amazon Web Services (AWS) released the SageMaker platform in 2017, which is a cloud machine-learning platform. SageMaker introduced a requester-friendly interface wrapping around MTurk called Ground Truth (GT) to design and generate HITs for recording worker responses that directly could be integrated with machine-learning algorithms.

Our Project

One of the growing problems due to the Internet age is misinformation. This happens when false information is shared with network circles through social media whether intentionally or unintentionally. Due to increasingly strong artificial intelligence, we now have the capability to create Deepfakes. Deepfakes are artificially created media (images or videos) that look like real humans. The Deepfake could be of a real person (as you can see in the Obama Deepfake image above), or it could just be a person who does not exist and is a mere combination of facial features from several real people. Deepfakes can be strong propagators of disinformation (intentional misinformation) because of how easy it is becoming to get access to a deep learning model that can generate Deepfakes.

There is currently ongoing research to create models that can detect Deepfakes including an ongoing challenge by Facebook AI to understand the progress on Deepfake detection technology that uses the Facebook DeepFake Detection Challenge Dataset (DFDC). However, the problem exists that training a model to be good would require large amounts of labeled data. We decided to tackle the question of whether the crowd is good at differentiating Deepfakes from real media, and especially does it make a difference if the media is an image versus a video.

We chose certain datasets to source Deepfake images, Deepfake videos, real images, and real videos. Funnily enough, we used the DFDC dataset to get the labeled video samples to understand whether humans are good at labeling these video samples correctly (a chicken and egg problem). We created HITs using SageMaker GT that thoroughly explained what exactly are Deepfakes and what we expect from the workers. We used past research knowledge to design the HITs to be somewhat resistant to malicious intentions while also providing us with good results to analyze.

We designed the HIT such that each worker would analyze two images and two videos (not necessarily one each of Deepfake and real media) by not only labeling whether a media is Deepfake or not, but also providing a reasoning of how they reached their judgment. Additionally, each worker also completed a survey that collected demographic information (age, race, gender) and cognitive level information (through CRT questions). As we were trying to understand if the crowd as a whole is capable of detecting Deepfakes and not necessarily individuals, each media was analyzed by five different workers.

Through this experiment, we wanted to analyze a few different aspects of human intelligence to label Deepfakes:

1. Does the crowd detect deepfaked videos better than deepfaked images?
2. Does the crowd reason differently for images and videos?
3. What features does the crowd use to aid in Deepfake detection?
4. What demographic of crowd has the highest accuracy?
5. Does performing well on a CRT test imply a higher accuracy at identifying Deepfakes?

Overall, we came to an understanding that the crowd is generally better at detecting deepfaked videos than images. We understood that this is likely because of motion information also being available when watching videos as compared to a still image. We discovered that generally for images, the crowd did not have any features that really stood out, but for videos, the motion features really helped differentiate Deepfakes from real media. Additionally, we were not able to get conclusive results based on the demographic features and CRT results because of our small sample size and heavily biased demographics (e.g. 93% of our samples were labeled by men).

So while it is technically possible to use a crowd to label Deepfake media (mostly videos), it is probably better with a much larger crowd and much higher redundancy, and with good guidelines outlining the target features to help distinguish Deepfakes from real media.

This is a concise overview of the results described in our paper, so you can read our paper to get more insight and explanation for our results!