Why my DeepFake detection work will stay away from Deep Learning
Recently, I picked up new research work in Deepfakes Detection. I am interested in this field for a number of reasons, a major one being that there is a lot of potential for exploration here. Unlike fields like Customer Segmentation, Health System Analysis, or Parkinson’s Disease detection (my prior work experience), the rules of engagement are messy.
With the ever-changing nature of DeepFakes, we don’t have a true representative dataset that I can just analyze.
People find new ways to create a DeepFake, which will throw off detection algorithms. The addition of more types of videos (especially as the world goes more digital) and formats further muddies the waters. Thus I need my research to be scalable to a large number of distributions (types of DeepFakes), even if they is created using new techniques. And I need to make sure that my solution is not uber expensive, otherwise, my solution would not be useful. So what does any of this have to do with avoiding DeepFakes? To understand that first, let’s talk about what my work is and how I see people using my solution.
What is my idea?
Watch the video on the left to understand the idea in depth (tl;dr in the next para if you’re busy). It is about the research I had done to validate my hunch. It’s my introduction to the related research and a very brief overview of the idea. I had originally used this compilation to give a talk at my university’s RIR (Research In-Ring) Talks, where people can present new ideas and findings. I got my current work as a direct consequence of that talk and my slide show compilation. The slides are linked here. I will update the slides as new information comes up.
For those of you that aren’t interested in watching the whole thing, I will give you a tl;dr. Simply put, my idea is to take an image and run all kinds of forensic analysis/feature extraction/downsampling algorithms on it. The output of this would be a set of images that are downsampled with certain features highlighted. We use these images as input to a classifier to identify if something is a possible DeepFake. Using the downsampled images is significantly cheaper (even if we have more to run) Skeptical if this would work? Well take a breather because
The artifacts can be detected. Therefore we should be able to use a combination of them to check if an image is a DeepFake or not. But now for the reason that you clicked this video: “Why the beef against Deep Learning?” Simply put, it’s too expensive to be scalable in a lot of contexts. Want more? Read on. We also have people achieving results with similar approaches such as:
Where will you see DeepFakes (most likely)?
Twitter, Facebook, other social media. All these sites operate at a scale of hundreds of millions of images EVERY DAY. Just FB has 250 million images uploaded daily.
If we run expensive DeepLearning algorithms on every image uploaded, we will crash servers very very quickly. While the 90%+ accuracy is important, our solutions need to be useful in the context within which they are going to applied, not an isolated lab/laptop/cloud. That’s where my target is. I intend to build a cost-effective filter. Something that can vet out a majority of images (so that we can run the more accurate, heavy detectors on the images that need it). It’s also important that we cut off the number of false negatives. The following video summarises this article (and why I will be focusing on cutting down false negatives). The video is a longer dive into the philosophy of my approach as well as some ideas I have to go after false negatives.
That’s about it from my side. As always, do share your thoughts in the comments below. Would love to get more perspectives on my research and ideas. Follow me here to keep up with my work. Share it with whoever might be interested. Below are all the places you could reach out to me. Peace.
Reach Out To Me
Please leave your feedback on this article below. If this was useful to you, please share it and follow me here.
Check out my other articles on Medium. : https://rb.gy/zn1aiu
My YouTube. It’s a work in progress haha: https://rb.gy/88iwdd
Reach out to me on LinkedIn. Let’s connect: https://rb.gy/m5ok2y
My Twitter: https://twitter.com/Machine01776819
My Substack: https://devanshacc.substack.com/
If you would like to work with me email me: firstname.lastname@example.org
Live conversations at twitch here: https://rb.gy/zlhk9y
To get updates on my content- Instagram: https://rb.gy/gmvuy9
Get a free stock on Robinhood: https://join.robinhood.com/fnud75