The kuhn-kloo-zhuhn
I know it’s been ages now since I wrote about the research progress, but what can I say, sometimes it’s just work.
No matter how much you love them, Writing this article has become work for me in the past month(#facts 🙄)
I had so much to do and couldn’t focus on a single bit. These are the moments i just can’t handle well. Maybe I should learn this year.
Anywho🥲
In the final stage of research & being so near to a goal, I got distracted, but here we are again. I finished the research (yuhoooooooooooo 🙌🏼), wrote the thesis & I am now trying to summarize it in the simplest way possible.
In the last phase of research, I took trained machine learning models EfficientNet-B4 & B7. Pre-trained models act differently and after every epoch of training, you get a new version of the same model.
Let’s assume you have EfficientNet-B4. Now you train it for 100 frames/data points, which means you have EfficientNet-B4–100Epoch as your machine learning model!!! AMAZING RIGHT?
In the same way, if you keep doing it for 1000 data points, then 10,000 data points….you will have EfficientNet-B4 with 10,000 layers. Each layer is a model itself.
Once I finished training EfficientNet-B4 & B7 on CelebDF (Li et al.,2019) and the DFDC (Facebook, 2020) alongside FaceForensics++ (Rössler et al., 2019) benchmark dataset of 1000 images & then measured their deepfake detection performance.
Numbers…. yeah yeah!😈
After getting the results, I used those numbers in McNemar’s test. Ofc! 🤓
In literature-y lingo, it means our p-value surpassed the assumed threshold (⍺=0.05), the study cannot reject the null hypothesis and assumes no significant difference in the effectiveness of EfficientNet-B4 & B7.
well..! well..! well..!
On the practical assessment of research, if models are trained correctly with advantageous dataset quality, the CNNs model can be combined to outperform the individual detection capabilities of EfficientNet-B4 and EfficientNet-B7. Furthermore, no significant differences were determined in the effectiveness of these techniques, which exist as a breakthrough. Therefore, both EfficientNet-B4 and EfficientNet-B7 can be reliable deepfake detection techniques for a given time.
Profound research in the scholarly world.🥳
don’t forget to appreciate your brain cells & little ride-along friend chatGPT!! by Jasmin Bharadiya