![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F493138%2Fb23eb618234c258ea484b115a8c31e55%2Fearliest_deepfake.gif?generation=1587540202680369&alt=media\")
Being curious about possible performance on organic videos, I took the oldest and the most recent videos from the Ctrl Shift Face YouTube channel.
\n\nThe performance of my model on the oldest video was far from perfect, but, broadly speaking, the model was able to distinguish an authentic face from the fake one. Here's one nice transition from authentic to fake:
\n\n
Yet, on the most recent deepfake video the model had no clue...\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F493138%2Ff258c9ebd27cbb8aa28e796a9d56ef7a%2Flatest_deepfake.gif?generation=1587540226571269&alt=media\")
My takeaways from this are:\n1. Likely, it's gonna be a never-ending arms race between deepfake creators and deepfake detectors. \n2. Initially I thought that I might avoid considerable shake-up, as my model is simple and I made 5 main submissions during the last 3 days of the competition, which should limit overfitting to the public LB. Now I believe I might experience an ugly shake-up, as I didn't include a single organic video into my training...
", "messages": [ { "id": 816357, "postDate": "2020-04-22T09:57:07.110Z", "content": "Being curious about possible performance on organic videos, I took the oldest and the most recent videos from the Ctrl Shift Face YouTube channel.
\n\nThe performance of my model on the oldest video was far from perfect, but, broadly speaking, the model was able to distinguish an authentic face from the fake one. Here's one nice transition from authentic to fake:
\n\n
Yet, on the most recent deepfake video the model had no clue...\n
My takeaways from this are:\n1. Likely, it's gonna be a never-ending arms race between deepfake creators and deepfake detectors. \n2. Initially I thought that I might avoid considerable shake-up, as my model is simple and I made 5 main submissions during the last 3 days of the competition, which should limit overfitting to the public LB. Now I believe I might experience an ugly shake-up, as I didn't include a single organic video into my training...
", "rawMarkdown": "Being curious about possible performance on organic videos, I took the oldest and the most recent videos from the [*Ctrl Shift Face* YouTube channel](https://www.youtube.com/channel/UCKpH0CKltc73e4wh0_pgL3g/videos). \n\nThe performance of my model on **the oldest video** was far from perfect, but, broadly speaking, the model was able to distinguish an authentic face from the fake one. Here's one nice transition from authentic to fake:\n\n\n\nYet, on **the most recent deepfake video** the model had no clue...\n\n\n**My takeaways from this are**:\n1. Likely, it's gonna be a never-ending arms race between deepfake creators and deepfake detectors. \n2. Initially I thought that I might avoid considerable shake-up, as my model is simple and I made 5 main submissions during the last 3 days of the competition, which should limit overfitting to the public LB. Now I believe I might experience an ugly shake-up, as I didn't include a single organic video into my training... \n", "votes": 9 } ], "comments": [], "raw_markdown_by_id": { "816357": "Being curious about possible performance on organic videos, I took the oldest and the most recent videos from the [*Ctrl Shift Face* YouTube channel](https://www.youtube.com/channel/UCKpH0CKltc73e4wh0_pgL3g/videos). \n\nThe performance of my model on **the oldest video** was far from perfect, but, broadly speaking, the model was able to distinguish an authentic face from the fake one. Here's one nice transition from authentic to fake:\n\n\n\nYet, on **the most recent deepfake video** the model had no clue...\n\n\n**My takeaways from this are**:\n1. Likely, it's gonna be a never-ending arms race between deepfake creators and deepfake detectors. \n2. Initially I thought that I might avoid considerable shake-up, as my model is simple and I made 5 main submissions during the last 3 days of the competition, which should limit overfitting to the public LB. Now I believe I might experience an ugly shake-up, as I didn't include a single organic video into my training... \n" } }