I want to share my potentially naive idea with people smarter than me. If there's something to it, maybe someone who knows what they're doing can use it.
\n\nWould it be useful to generate a deepfake (images or videos) from the incoming video (or a portion of it) (call it video A) as a feature in your model? I wonder if the resulting video/image might be telling.\n1. Would it have more exaggerated artifacts if the target was a deepfake?\n2. Could you compare it to the incoming video (video A) and use the differences/similarities as features?\n3. Could you derive other useful features from it?
\n\nAs I said, I am a ML noob so if anyone has some extra time to instruct me on the problems with this approach, it would be very much appreciated!
", "messages": [ { "id": "737187", "postDate": "02/05/2020 03:03:02", "content": "I want to share my potentially naive idea with people smarter than me. If there's something to it, maybe someone who knows what they're doing can use it.
\n\nWould it be useful to generate a deepfake (images or videos) from the incoming video (or a portion of it) (call it video A) as a feature in your model? I wonder if the resulting video/image might be telling.\n1. Would it have more exaggerated artifacts if the target was a deepfake?\n2. Could you compare it to the incoming video (video A) and use the differences/similarities as features?\n3. Could you derive other useful features from it?
\n\nAs I said, I am a ML noob so if anyone has some extra time to instruct me on the problems with this approach, it would be very much appreciated!
", "rawMarkdown": "I want to share my potentially naive idea with people smarter than me. If there's something to it, maybe someone who knows what they're doing can use it.\n\nWould it be useful to generate a deepfake (images or videos) from the incoming video (or a portion of it) (call it video A) as a feature in your model? I wonder if the resulting video/image might be telling.\n1. Would it have more exaggerated artifacts if the target was a deepfake?\n2. Could you compare it to the incoming video (video A) and use the differences/similarities as features?\n3. Could you derive other useful features from it?\n\nAs I said, I am a ML noob so if anyone has some extra time to instruct me on the problems with this approach, it would be very much appreciated!", "votes": null }, { "id": "737197", "postDate": "02/05/2020 03:30:08", "content": "For number 2, it is not possible because kaggle will not provide what is the original video during submission(or otherwise it will be the biggest data leak in history).
", "rawMarkdown": "For number 2, it is not possible because kaggle will not provide what is the original video during submission(or otherwise it will be the biggest data leak in history).", "votes": null }, { "id": "737204", "postDate": "02/05/2020 03:38:54", "content": "For #2, I meant compare the new deepfake to the video your net made the deepfake from.
", "rawMarkdown": "For #2, I meant compare the new deepfake to the video your net made the deepfake from.", "votes": null }, { "id": "737206", "postDate": "02/05/2020 03:41:13", "content": "What do you mean by generate deep fake videos? to add augmentations to it or to train a GAN or something.
", "rawMarkdown": "What do you mean by generate deep fake videos? to add augmentations to it or to train a GAN or something.", "votes": null }, { "id": "737634", "postDate": "02/05/2020 15:53:51", "content": "Irrespective, unless the newly generated deepfake is within the same distribution as in the train data, it'll only add to the mess.
", "rawMarkdown": "Irrespective, unless the newly generated deepfake is within the same distribution as in the train data, it'll only add to the mess.", "votes": null }, { "id": "737721", "postDate": "02/05/2020 17:28:18", "content": "@unkownhihi I guess I mean use a GAN with the (potentially) altered video as input, along with the face that's in it. Would the resulting video be telling at all?\nFor instance, if the original was a fake, after performing this operation its history would look like this:
\n\na real video + a new face => encoded/decoded => a deepfake => encoded/decoded => a new deepfake
\n\nAnd if the original was not a fake, after performing this operation its history would look like this:
\n\na real video => encoded/decoded => a deepfake
\n\nObviously you won't know the history, but the question is whether a video that's gone through two encoding/decoding cycles looks a lot different than a video that's only gone through one.
\n\nThe thought comes from knowing what happens when you take an image or video and compress it, decompress it, and then compress it again. It loses more quality than if you had only compressed it once (far as I know).
", "rawMarkdown": "unkownhihi I guess I mean use a GAN with the (potentially) altered video as input, along with the face that's in it. Would the resulting video be telling at all?\nFor instance, if the original was a fake, after performing this operation its history would look like this:\n\na real video + a new face => encoded/decoded => a deepfake => encoded/decoded => a new deepfake\n\nAnd if the original was not a fake, after performing this operation its history would look like this:\n\na real video => encoded/decoded => a deepfake\n\nObviously you won't know the history, but the question is whether a video that's gone through two encoding/decoding cycles looks a lot different than a video that's only gone through one.\n\nThe thought comes from knowing what happens when you take an image or video and compress it, decompress it, and then compress it again. It loses more quality than if you had only compressed it once (far as I know).", "votes": null }, { "id": "737743", "postDate": "02/05/2020 17:59:10", "content": "@akashnandi yeah I guess that makes sense. However, maybe you could establish a baseline of what your GAN typically produces from original, real videos, and what it produces when the video is not real. Thanks for the feedback.
", "rawMarkdown": "akashnandi yeah I guess that makes sense. However, maybe you could establish a baseline of what **your** GAN typically produces from original, real videos, and what it produces when the video is not real. Thanks for the feedback.", "votes": null } ], "comments": [ { "id": 737197, "author_name": "unkownhihi", "author_url": "", "post_date": "02/05/2020 03:30:08", "content": "For number 2, it is not possible because kaggle will not provide what is the original video during submission(or otherwise it will be the biggest data leak in history).
", "votes": null, "replies": [ { "id": 737204, "author_name": "n1njaj3f", "author_url": "", "post_date": "02/05/2020 03:38:54", "content": "For #2, I meant compare the new deepfake to the video your net made the deepfake from.
", "votes": null, "replies": [] }, { "id": 737206, "author_name": "unkownhihi", "author_url": "", "post_date": "02/05/2020 03:41:13", "content": "What do you mean by generate deep fake videos? to add augmentations to it or to train a GAN or something.
", "votes": null, "replies": [] }, { "id": 737634, "author_name": "akashnandi", "author_url": "", "post_date": "02/05/2020 15:53:51", "content": "Irrespective, unless the newly generated deepfake is within the same distribution as in the train data, it'll only add to the mess.
", "votes": null, "replies": [] }, { "id": 737721, "author_name": "n1njaj3f", "author_url": "", "post_date": "02/05/2020 17:28:18", "content": "@unkownhihi I guess I mean use a GAN with the (potentially) altered video as input, along with the face that's in it. Would the resulting video be telling at all?\nFor instance, if the original was a fake, after performing this operation its history would look like this:
\n\na real video + a new face => encoded/decoded => a deepfake => encoded/decoded => a new deepfake
\n\nAnd if the original was not a fake, after performing this operation its history would look like this:
\n\na real video => encoded/decoded => a deepfake
\n\nObviously you won't know the history, but the question is whether a video that's gone through two encoding/decoding cycles looks a lot different than a video that's only gone through one.
\n\nThe thought comes from knowing what happens when you take an image or video and compress it, decompress it, and then compress it again. It loses more quality than if you had only compressed it once (far as I know).
", "votes": null, "replies": [] }, { "id": 737743, "author_name": "n1njaj3f", "author_url": "", "post_date": "02/05/2020 17:59:10", "content": "@akashnandi yeah I guess that makes sense. However, maybe you could establish a baseline of what your GAN typically produces from original, real videos, and what it produces when the video is not real. Thanks for the feedback.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "737187": "I want to share my potentially naive idea with people smarter than me. If there's something to it, maybe someone who knows what they're doing can use it.\n\nWould it be useful to generate a deepfake (images or videos) from the incoming video (or a portion of it) (call it video A) as a feature in your model? I wonder if the resulting video/image might be telling.\n1. Would it have more exaggerated artifacts if the target was a deepfake?\n2. Could you compare it to the incoming video (video A) and use the differences/similarities as features?\n3. Could you derive other useful features from it?\n\nAs I said, I am a ML noob so if anyone has some extra time to instruct me on the problems with this approach, it would be very much appreciated!", "737197": "For number 2, it is not possible because kaggle will not provide what is the original video during submission(or otherwise it will be the biggest data leak in history).", "737204": "For #2, I meant compare the new deepfake to the video your net made the deepfake from.", "737206": "What do you mean by generate deep fake videos? to add augmentations to it or to train a GAN or something.", "737634": "Irrespective, unless the newly generated deepfake is within the same distribution as in the train data, it'll only add to the mess.", "737721": "unkownhihi I guess I mean use a GAN with the (potentially) altered video as input, along with the face that's in it. Would the resulting video be telling at all?\nFor instance, if the original was a fake, after performing this operation its history would look like this:\n\na real video + a new face => encoded/decoded => a deepfake => encoded/decoded => a new deepfake\n\nAnd if the original was not a fake, after performing this operation its history would look like this:\n\na real video => encoded/decoded => a deepfake\n\nObviously you won't know the history, but the question is whether a video that's gone through two encoding/decoding cycles looks a lot different than a video that's only gone through one.\n\nThe thought comes from knowing what happens when you take an image or video and compress it, decompress it, and then compress it again. It loses more quality than if you had only compressed it once (far as I know).", "737743": "akashnandi yeah I guess that makes sense. However, maybe you could establish a baseline of what **your** GAN typically produces from original, real videos, and what it produces when the video is not real. Thanks for the feedback." }, "source": "meta" }