Couldn't you use a great discriminator to build a much better generator? GANs work this way...
", "messages": [ { "id": "692884", "postDate": "12/11/2019 20:17:03", "content": "Couldn't you use a great discriminator to build a much better generator? GANs work this way...
", "rawMarkdown": "Couldn't you use a great discriminator to build a much better generator? GANs work this way...", "votes": null }, { "id": "692938", "postDate": "12/11/2019 21:33:08", "content": "You win the prize for discovering the true objective of this competition!
", "rawMarkdown": "You win the prize for discovering the true objective of this competition!", "votes": null }, { "id": "693016", "postDate": "12/12/2019 00:42:13", "content": "Exactly my thoughts...
", "rawMarkdown": "Exactly my thoughts...", "votes": null }, { "id": "693188", "postDate": "12/12/2019 05:50:25", "content": "Usually separate detectors are better than discriminators trained in a GAN. This also usually implies that the discriminator would probably be too strong for GAN training.
\n\nThere's a chance though.
", "rawMarkdown": "Usually separate detectors are better than discriminators trained in a GAN. This also usually implies that the discriminator would probably be too strong for GAN training.\n\nThere's a chance though.", "votes": null }, { "id": "693449", "postDate": "12/12/2019 11:36:03", "content": "Let's say you had a discriminator that is 100% correct at all times, wouldn't it be possible to artificially make it worse so that at beginning of training it keeps pace with the generator? As the generator gets better, you can slowly ramp up the quality of the discriminator as well.
", "rawMarkdown": "Let's say you had a discriminator that is 100% correct at all times, wouldn't it be possible to artificially make it worse so that at beginning of training it keeps pace with the generator? As the generator gets better, you can slowly ramp up the quality of the discriminator as well.", "votes": null }, { "id": "693464", "postDate": "12/12/2019 11:53:52", "content": "It's an endless \"arms race\" that spans not only this specific instance of deepfake but rather any competition in industry (and nature as well). There's always an \"attacking\" agent (technology/predator) and \"defending\" agent. Whenever the attackers come up with new methods/technology/capabilities the \"defenders\" respond with improving their capabilities.\nTherefore, I don't think this competition (as successful as it may turn out) will be the end of this \"arms race\" but rather \"another brick in the wall\". I foresee a reaction by developers of GANs and other video/voice manipulation technologies to come up with even better ways to outsmart the winners of this competition. \nSo brace yourself, this is going to be a long journey and I expect we shall meet in similar competitions/challenges many more times in the future.
", "rawMarkdown": "It's an endless \"arms race\" that spans not only this specific instance of deepfake but rather any competition in industry (and nature as well). There's always an \"attacking\" agent (technology/predator) and \"defending\" agent. Whenever the attackers come up with new methods/technology/capabilities the \"defenders\" respond with improving their capabilities.\nTherefore, I don't think this competition (as successful as it may turn out) will be the end of this \"arms race\" but rather \"another brick in the wall\". I foresee a reaction by developers of GANs and other video/voice manipulation technologies to come up with even better ways to outsmart the winners of this competition. \nSo brace yourself, this is going to be a long journey and I expect we shall meet in similar competitions/challenges many more times in the future.", "votes": null }, { "id": "693602", "postDate": "12/12/2019 15:00:07", "content": "I was just thinking about this... Today when people see deep fakes they can tell something is not right. A person may not know exactly what is not right about the image or sound, but they notice a difference. As we get better at detecting the fakes, the fakes will improve to the point the humans will no longer be able to tell the difference. It seems like this is already out of control.
", "rawMarkdown": "I was just thinking about this... Today when people see deep fakes they can tell something is not right. A person may not know exactly what is not right about the image or sound, but they notice a difference. As we get better at detecting the fakes, the fakes will improve to the point the humans will no longer be able to tell the difference. It seems like this is already out of control.", "votes": null }, { "id": "694281", "postDate": "12/13/2019 11:40:40", "content": "My two cents on this topic: Yes this competition might help in creating better deep fakes, unless the winner provably discriminates some ground truth with any tampered videos (instead of just finding some gaps in attempts).
\n\nI believe the path to prevent deep fakes and identify deep fakes lies in cryptography, specifically in encoding the video information with creator's signature embedded in the video. Any attempt to tamper with the video would break the signatures encoding/verification and hence show that it is fake. This could become a embedded in/part of the hardwares/softwares like cameras etc.
", "rawMarkdown": "My two cents on this topic: Yes this competition might help in creating better deep fakes, unless the winner provably discriminates some ground truth with any tampered videos (instead of just finding some gaps in attempts).\n\nI believe the path to prevent deep fakes and identify deep fakes lies in cryptography, specifically in encoding the video information with creator's signature embedded in the video. Any attempt to tamper with the video would break the signatures encoding/verification and hence show that it is fake. This could become a embedded in/part of the hardwares/softwares like cameras etc.", "votes": null }, { "id": "695290", "postDate": "12/14/2019 22:41:56", "content": "virus vs. antivirus...\ndoping vs. antidoping...\nit's a known market dynamic...
", "rawMarkdown": "virus vs. antivirus...\ndoping vs. antidoping...\nit's a known market dynamic...", "votes": null }, { "id": "733578", "postDate": "01/31/2020 10:14:50", "content": "I think FB has two goals here.
\n\nTheir short term goal is to combat deepfake attempts to influence the 2020 US election. Even if it leads to better deepfakes in 2021, if they can keep control for 2020 they won't have to deal with another presidential election until 2024. They already have their own deepfake detection algorithms. The timing works out so that if the winner of this comp comes up with something even better, or just something new that can ensemble with their current approach, FB may have time to add it to their pipeline before the election really heats up. (Hopefully - obviously Kaggle code won't directly work in production.)
\n\nTheir long term strategy is to provide the DFDC dataset for deepfake detection research in general, essentially looking to spur the overall defensive efforts instead of just building one new weapon in the metaphorical arms race.
", "rawMarkdown": "I think FB has two goals here. \n\nTheir short term goal is to combat deepfake attempts to influence the 2020 US election. Even if it leads to better deepfakes in 2021, if they can keep control for 2020 they won't have to deal with another presidential election until 2024. They already have their own deepfake detection algorithms. The timing works out so that if the winner of this comp comes up with something even better, or just something new that can ensemble with their current approach, FB may have time to add it to their pipeline before the election really heats up. (Hopefully - obviously Kaggle code won't directly work in production.) \n\nTheir long term strategy is to provide the DFDC dataset for deepfake detection research in general, essentially looking to spur the overall defensive efforts instead of just building one new weapon in the metaphorical arms race.", "votes": null }, { "id": "756558", "postDate": "02/25/2020 20:49:44", "content": "That's a very interesting approach and solution to the current problem. Any links to such attempts to implement these ideas? POCs or half-backed ideas are fine to share of course. Thanks!
", "rawMarkdown": "That's a very interesting approach and solution to the current problem. Any links to such attempts to implement these ideas? POCs or half-backed ideas are fine to share of course. Thanks!", "votes": null } ], "comments": [ { "id": 692938, "author_name": "interneuron", "author_url": "", "post_date": "12/11/2019 21:33:08", "content": "You win the prize for discovering the true objective of this competition!
", "votes": null, "replies": [] }, { "id": 693016, "author_name": "ma7555", "author_url": "", "post_date": "12/12/2019 00:42:13", "content": "Exactly my thoughts...
", "votes": null, "replies": [] }, { "id": 693188, "author_name": "jesperdramsch", "author_url": "", "post_date": "12/12/2019 05:50:25", "content": "Usually separate detectors are better than discriminators trained in a GAN. This also usually implies that the discriminator would probably be too strong for GAN training.
\n\nThere's a chance though.
", "votes": null, "replies": [ { "id": 693449, "author_name": "humananalog", "author_url": "", "post_date": "12/12/2019 11:36:03", "content": "Let's say you had a discriminator that is 100% correct at all times, wouldn't it be possible to artificially make it worse so that at beginning of training it keeps pace with the generator? As the generator gets better, you can slowly ramp up the quality of the discriminator as well.
", "votes": null, "replies": [] } ] }, { "id": 693464, "author_name": "shaibagon", "author_url": "", "post_date": "12/12/2019 11:53:52", "content": "It's an endless \"arms race\" that spans not only this specific instance of deepfake but rather any competition in industry (and nature as well). There's always an \"attacking\" agent (technology/predator) and \"defending\" agent. Whenever the attackers come up with new methods/technology/capabilities the \"defenders\" respond with improving their capabilities.\nTherefore, I don't think this competition (as successful as it may turn out) will be the end of this \"arms race\" but rather \"another brick in the wall\". I foresee a reaction by developers of GANs and other video/voice manipulation technologies to come up with even better ways to outsmart the winners of this competition. \nSo brace yourself, this is going to be a long journey and I expect we shall meet in similar competitions/challenges many more times in the future.
", "votes": null, "replies": [] }, { "id": 693602, "author_name": "byronkd", "author_url": "", "post_date": "12/12/2019 15:00:07", "content": "I was just thinking about this... Today when people see deep fakes they can tell something is not right. A person may not know exactly what is not right about the image or sound, but they notice a difference. As we get better at detecting the fakes, the fakes will improve to the point the humans will no longer be able to tell the difference. It seems like this is already out of control.
", "votes": null, "replies": [] }, { "id": 694281, "author_name": "pyrole", "author_url": "", "post_date": "12/13/2019 11:40:40", "content": "My two cents on this topic: Yes this competition might help in creating better deep fakes, unless the winner provably discriminates some ground truth with any tampered videos (instead of just finding some gaps in attempts).
\n\nI believe the path to prevent deep fakes and identify deep fakes lies in cryptography, specifically in encoding the video information with creator's signature embedded in the video. Any attempt to tamper with the video would break the signatures encoding/verification and hence show that it is fake. This could become a embedded in/part of the hardwares/softwares like cameras etc.
", "votes": null, "replies": [ { "id": 756558, "author_name": "yassinealouini", "author_url": "", "post_date": "02/25/2020 20:49:44", "content": "That's a very interesting approach and solution to the current problem. Any links to such attempts to implement these ideas? POCs or half-backed ideas are fine to share of course. Thanks!
", "votes": null, "replies": [] } ] }, { "id": 695290, "author_name": "carlossouza", "author_url": "", "post_date": "12/14/2019 22:41:56", "content": "virus vs. antivirus...\ndoping vs. antidoping...\nit's a known market dynamic...
", "votes": null, "replies": [] }, { "id": 733578, "author_name": "jamesthornton", "author_url": "", "post_date": "01/31/2020 10:14:50", "content": "I think FB has two goals here.
\n\nTheir short term goal is to combat deepfake attempts to influence the 2020 US election. Even if it leads to better deepfakes in 2021, if they can keep control for 2020 they won't have to deal with another presidential election until 2024. They already have their own deepfake detection algorithms. The timing works out so that if the winner of this comp comes up with something even better, or just something new that can ensemble with their current approach, FB may have time to add it to their pipeline before the election really heats up. (Hopefully - obviously Kaggle code won't directly work in production.)
\n\nTheir long term strategy is to provide the DFDC dataset for deepfake detection research in general, essentially looking to spur the overall defensive efforts instead of just building one new weapon in the metaphorical arms race.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "692884": "Couldn't you use a great discriminator to build a much better generator? GANs work this way...", "692938": "You win the prize for discovering the true objective of this competition!", "693016": "Exactly my thoughts...", "693188": "Usually separate detectors are better than discriminators trained in a GAN. This also usually implies that the discriminator would probably be too strong for GAN training.\n\nThere's a chance though.", "693449": "Let's say you had a discriminator that is 100% correct at all times, wouldn't it be possible to artificially make it worse so that at beginning of training it keeps pace with the generator? As the generator gets better, you can slowly ramp up the quality of the discriminator as well.", "693464": "It's an endless \"arms race\" that spans not only this specific instance of deepfake but rather any competition in industry (and nature as well). There's always an \"attacking\" agent (technology/predator) and \"defending\" agent. Whenever the attackers come up with new methods/technology/capabilities the \"defenders\" respond with improving their capabilities.\nTherefore, I don't think this competition (as successful as it may turn out) will be the end of this \"arms race\" but rather \"another brick in the wall\". I foresee a reaction by developers of GANs and other video/voice manipulation technologies to come up with even better ways to outsmart the winners of this competition. \nSo brace yourself, this is going to be a long journey and I expect we shall meet in similar competitions/challenges many more times in the future.", "693602": "I was just thinking about this... Today when people see deep fakes they can tell something is not right. A person may not know exactly what is not right about the image or sound, but they notice a difference. As we get better at detecting the fakes, the fakes will improve to the point the humans will no longer be able to tell the difference. It seems like this is already out of control.", "694281": "My two cents on this topic: Yes this competition might help in creating better deep fakes, unless the winner provably discriminates some ground truth with any tampered videos (instead of just finding some gaps in attempts).\n\nI believe the path to prevent deep fakes and identify deep fakes lies in cryptography, specifically in encoding the video information with creator's signature embedded in the video. Any attempt to tamper with the video would break the signatures encoding/verification and hence show that it is fake. This could become a embedded in/part of the hardwares/softwares like cameras etc.", "695290": "virus vs. antivirus...\ndoping vs. antidoping...\nit's a known market dynamic...", "733578": "I think FB has two goals here. \n\nTheir short term goal is to combat deepfake attempts to influence the 2020 US election. Even if it leads to better deepfakes in 2021, if they can keep control for 2020 they won't have to deal with another presidential election until 2024. They already have their own deepfake detection algorithms. The timing works out so that if the winner of this comp comes up with something even better, or just something new that can ensemble with their current approach, FB may have time to add it to their pipeline before the election really heats up. (Hopefully - obviously Kaggle code won't directly work in production.) \n\nTheir long term strategy is to provide the DFDC dataset for deepfake detection research in general, essentially looking to spur the overall defensive efforts instead of just building one new weapon in the metaphorical arms race.", "756558": "That's a very interesting approach and solution to the current problem. Any links to such attempts to implement these ideas? POCs or half-backed ideas are fine to share of course. Thanks!" }, "source": "meta" }