DeepFakeAudio

Audio:Real || Fake

Inspiration

We had heard about a story where a lad got a call that was deep faked, and using it, they were scammed over $250,000. We thought we would try and help others from suffering the same fate, as that can completely ruin someones life. Not everyone has large amounts of savings or spare capital, and losing that amount could be unrecoverable.

What it does

It takes and audio input and analyses it to detect if it is authentic or fake.

How we built it

We built it by training an AI based with a set of training data that had fake and authentic audio. After being trained on that, we had it predict if other audio was authentic or falsified.

Challenges we ran into

We ran into an issue of finding the data set, as well as downloading it in a timely manner. Much of the data that was easily available was suboptimal or unlabeled, as such making it very difficult to train with. The data availability is limiting, as you cannot test the model with the data that it is trained with, as the AI would already be familiar with the data. The next challenge outside of the data was the best algorithm to use for the analysis, as there are a variety of options, but we were looking for the one with the most accuracy. The GUI we were creating was also a challenge, as we were trying to think of the best way to display the information in an intuitive manner, when the majority of the program is not visually based.

Accomplishments that we're proud of

We are proud of making the model for the AI ourselves, rather than taking an already available one, as well as the high accuracy that the model has.

What we learned

We learned a fair bit with regards to font end development (GUI), as none of our team members had that as an area of greater knowledge and experience. There was also learning experience with AI, as not everyone was overly familiar with it in the areas from algorithm selection to model creation and implementation.

What's next for DeepFakeAudio

Next we would love to expand the training database to increase the knowledge and accuracy of the model. With that we would love to have it also be able to analysis the audio for video, so to try and see if videos are accurate. Many times the audio and lip movement on a video are delayed, so the ability to detect if the audio is authentic when one cannot align the lip movement (or see the lip movement) would be a great expansion. Ideally, we would also love to expand the detection to deep fake for video and images not just audio. Deep fake video can be just as harmful as audio, if not more so.