Recently Andrew Trask commented on twitter:
\n\n\n\n\n\n\nDetecting AI generated DeepFakes using AI is a losing game. I give it 2 years tops\n Eventually it will be impossible for the same reason it’s impossible to humans. It’s an information bottleneck.\n Example: “010010” <- is this fake or was this created based on real world events?
\n \nInstead a digital object (image, video, audio, etc.) should only be deemed “real” or “fake” based on which real people claim it to be so.\n We need the ability for everyday people to cryptographically sign content they create (tweets, photos, videos, phone calls, etc)
\n
What do you folks think about this comment? Do reply below
", "messages": [ { "id": "702938", "postDate": "12/25/2019 11:34:55", "content": "Recently Andrew Trask commented on twitter:
\n\n\n\n\n\n\nDetecting AI generated DeepFakes using AI is a losing game. I give it 2 years tops\n Eventually it will be impossible for the same reason it’s impossible to humans. It’s an information bottleneck.\n Example: “010010” <- is this fake or was this created based on real world events?
\n \nInstead a digital object (image, video, audio, etc.) should only be deemed “real” or “fake” based on which real people claim it to be so.\n We need the ability for everyday people to cryptographically sign content they create (tweets, photos, videos, phone calls, etc)
\n
What do you folks think about this comment? Do reply below
", "rawMarkdown": "Recently Andrew Trask commented on twitter:\n> Detecting AI generated DeepFakes using AI is a losing game. I give it 2 years tops\nEventually it will be impossible for the same reason it’s impossible to humans. It’s an information bottleneck.\nExample: “010010” <- is this fake or was this created based on real world events?\n\n> Instead a digital object (image, video, audio, etc.) should only be deemed “real” or “fake” based on which real people claim it to be so.\nWe need the ability for everyday people to cryptographically sign content they create (tweets, photos, videos, phone calls, etc)\n\n[source](https://twitter.com/iamtrask/status/1209563964377616387)\n\nWhat do you folks think about this comment? Do reply below", "votes": null }, { "id": "704081", "postDate": "12/27/2019 02:50:49", "content": "it's not enough but it's not loosing game, it's something that should be done along side many other things.. cryptographic signatures are good but it needs the media platforms to use them to be useful.. if you go to a bad website and use that as your source (like a lot of people do) it won't make a difference because you decided to trust that source, it's about how much the receivers care about it more than how authentic it is imo
", "rawMarkdown": "it's not enough but it's not loosing game, it's something that should be done along side many other things.. cryptographic signatures are good but it needs the media platforms to use them to be useful.. if you go to a bad website and use that as your source (like a lot of people do) it won't make a difference because you decided to trust that source, it's about how much the receivers care about it more than how authentic it is imo", "votes": null }, { "id": "705540", "postDate": "12/29/2019 02:45:47", "content": "I don't think it will be impossible for computers and I don't believe the two year timeline.
", "rawMarkdown": "I don't think it will be impossible for computers and I don't believe the two year timeline.", "votes": null }, { "id": "705602", "postDate": "12/29/2019 05:57:59", "content": "I think he's right. As we learn how to beat a given level of deep fake technology, and the method of detection becomes known, the deepfake creation technology will evolve to fool it. Eventually, the fakes become, basically, perfect. I think that the end result will be that videos are no longer useable evidence - a pretty catastrophic endpoint.
", "rawMarkdown": "I think he's right. As we learn how to beat a given level of deep fake technology, and the method of detection becomes known, the deepfake creation technology will evolve to fool it. Eventually, the fakes become, basically, perfect. I think that the end result will be that videos are no longer useable evidence - a pretty catastrophic endpoint.", "votes": null }, { "id": "705603", "postDate": "12/29/2019 06:04:13", "content": "I agree with him. Its like a classic GAN situation, all this will do is create a better discriminator for the Deep Generative Models to beat. At the end, its at the source of data where authentication can be managed.
", "rawMarkdown": "I agree with him. Its like a classic GAN situation, all this will do is create a better discriminator for the Deep Generative Models to beat. At the end, its at the source of data where authentication can be managed.", "votes": null }, { "id": "706743", "postDate": "12/30/2019 19:03:09", "content": "Well, spotting \"an information bottleneck\" here is very theoretical. There was a time where classifying cats vs dogs was considered impossible using NP completeness arguments.
", "rawMarkdown": "Well, spotting \"an information bottleneck\" here is very theoretical. There was a time where classifying cats vs dogs was considered impossible using NP completeness arguments.", "votes": null }, { "id": "707651", "postDate": "01/01/2020 08:26:33", "content": "Was classifying Cats vs Dogs so hard?
", "rawMarkdown": "Was classifying Cats vs Dogs so hard?", "votes": null }, { "id": "709747", "postDate": "01/03/2020 22:20:57", "content": "Intellectuals have their opinions for sure, but the industy dominates.
", "rawMarkdown": "Intellectuals have their opinions for sure, but the industy dominates.", "votes": null }, { "id": "713921", "postDate": "01/08/2020 19:58:24", "content": "As long as humans can identify a deepfake by artifacts and such, we should definitely be able to identify by the current line of works. I think people will more follow Andrew's advice once it reaches that tipping point. In any case, this depends on the context of where we want to identify DeepFakes. Anonymous videos on Reddit arent gonna be verified by any means. So, AI will be useful there even if the videos can fool humans.
", "rawMarkdown": "As long as humans can identify a deepfake by artifacts and such, we should definitely be able to identify by the current line of works. I think people will more follow Andrew's advice once it reaches that tipping point. In any case, this depends on the context of where we want to identify DeepFakes. Anonymous videos on Reddit arent gonna be verified by any means. So, AI will be useful there even if the videos can fool humans.", "votes": null }, { "id": "729019", "postDate": "01/25/2020 15:32:40", "content": "How will 'real people' claim that a digital object is real or fake?
\n\nThere were two recent incidents where experts couldn't determine if the video was real or fake - one of them caused a military coup (... a failed one though!) and the other a minister to resign.
\n\nDigital signature for verifying/ authenticating a digital object is a great idea. But it's more from the copyright angle.
\n\nDue to their far-reaching impact, DeepFakes are a social concern. And tools should be easily available to identify them. In fact we can equate it to the arms race in the Anti-Virus/Malware industry.
", "rawMarkdown": "How will *'real people'* claim that a digital object is real or fake? \n\nThere were two recent incidents where experts couldn't determine if the video was real or fake - one of them caused a [military coup (... a failed one though!)](https://www.motherjones.com/politics/2019/03/deepfake-gabon-ali-bongo) and the other a [minister to resign.](https://www.scmp.com/week-asia/politics/article/3014162/malaysian-minister-implicated-gay-sex-video-prompting) \n\nDigital signature for verifying/ authenticating a digital object is a great idea. But it's more from the copyright angle.\n\nDue to their far-reaching impact, DeepFakes are a social concern. And tools should be easily available to identify them. In fact we can equate it to the arms race in the Anti-Virus/Malware industry.", "votes": null }, { "id": "1848481", "postDate": "07/08/2022 16:51:40", "content": "I agree with him. It is necessary to to sign content they created, It can embed some unique Identifier verified my profiles ids, only those IDs which created the digital objects, or upload the object, date ob digital object creation should embed in the data, It will be helpful to categories the resource as a new creation or old. At the end, its at the source of data where authentication can be managed.
", "rawMarkdown": "I agree with him. It is necessary to to sign content they created, It can embed some unique Identifier verified my profiles ids, only those IDs which created the digital objects, or upload the object, date ob digital object creation should embed in the data, It will be helpful to categories the resource as a new creation or old. At the end, its at the source of data where authentication can be managed.", "votes": null } ], "comments": [ { "id": 1848481, "author_name": "muhammadirfanmalik", "author_url": "", "post_date": "07/08/2022 16:51:40", "content": "I agree with him. It is necessary to to sign content they created, It can embed some unique Identifier verified my profiles ids, only those IDs which created the digital objects, or upload the object, date ob digital object creation should embed in the data, It will be helpful to categories the resource as a new creation or old. At the end, its at the source of data where authentication can be managed.
", "votes": null, "replies": [] }, { "id": 704081, "author_name": "basharallabadi", "author_url": "", "post_date": "12/27/2019 02:50:49", "content": "it's not enough but it's not loosing game, it's something that should be done along side many other things.. cryptographic signatures are good but it needs the media platforms to use them to be useful.. if you go to a bad website and use that as your source (like a lot of people do) it won't make a difference because you decided to trust that source, it's about how much the receivers care about it more than how authentic it is imo
", "votes": null, "replies": [] }, { "id": 705540, "author_name": "drsilicon", "author_url": "", "post_date": "12/29/2019 02:45:47", "content": "I don't think it will be impossible for computers and I don't believe the two year timeline.
", "votes": null, "replies": [] }, { "id": 705602, "author_name": "petewills", "author_url": "", "post_date": "12/29/2019 05:57:59", "content": "I think he's right. As we learn how to beat a given level of deep fake technology, and the method of detection becomes known, the deepfake creation technology will evolve to fool it. Eventually, the fakes become, basically, perfect. I think that the end result will be that videos are no longer useable evidence - a pretty catastrophic endpoint.
", "votes": null, "replies": [] }, { "id": 705603, "author_name": "bismay", "author_url": "", "post_date": "12/29/2019 06:04:13", "content": "I agree with him. Its like a classic GAN situation, all this will do is create a better discriminator for the Deep Generative Models to beat. At the end, its at the source of data where authentication can be managed.
", "votes": null, "replies": [] }, { "id": 706743, "author_name": "alexis1024", "author_url": "", "post_date": "12/30/2019 19:03:09", "content": "Well, spotting \"an information bottleneck\" here is very theoretical. There was a time where classifying cats vs dogs was considered impossible using NP completeness arguments.
", "votes": null, "replies": [ { "id": 707651, "author_name": "kurianbenoy", "author_url": "", "post_date": "01/01/2020 08:26:33", "content": "Was classifying Cats vs Dogs so hard?
", "votes": null, "replies": [] } ] }, { "id": 709747, "author_name": "", "author_url": "", "post_date": "01/03/2020 22:20:57", "content": "Intellectuals have their opinions for sure, but the industy dominates.
", "votes": null, "replies": [] }, { "id": 713921, "author_name": "skrish13", "author_url": "", "post_date": "01/08/2020 19:58:24", "content": "As long as humans can identify a deepfake by artifacts and such, we should definitely be able to identify by the current line of works. I think people will more follow Andrew's advice once it reaches that tipping point. In any case, this depends on the context of where we want to identify DeepFakes. Anonymous videos on Reddit arent gonna be verified by any means. So, AI will be useful there even if the videos can fool humans.
", "votes": null, "replies": [] }, { "id": 729019, "author_name": "skylord", "author_url": "", "post_date": "01/25/2020 15:32:40", "content": "How will 'real people' claim that a digital object is real or fake?
\n\nThere were two recent incidents where experts couldn't determine if the video was real or fake - one of them caused a military coup (... a failed one though!) and the other a minister to resign.
\n\nDigital signature for verifying/ authenticating a digital object is a great idea. But it's more from the copyright angle.
\n\nDue to their far-reaching impact, DeepFakes are a social concern. And tools should be easily available to identify them. In fact we can equate it to the arms race in the Anti-Virus/Malware industry.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "702938": "Recently Andrew Trask commented on twitter:\n> Detecting AI generated DeepFakes using AI is a losing game. I give it 2 years tops\nEventually it will be impossible for the same reason it’s impossible to humans. It’s an information bottleneck.\nExample: “010010” <- is this fake or was this created based on real world events?\n\n> Instead a digital object (image, video, audio, etc.) should only be deemed “real” or “fake” based on which real people claim it to be so.\nWe need the ability for everyday people to cryptographically sign content they create (tweets, photos, videos, phone calls, etc)\n\n[source](https://twitter.com/iamtrask/status/1209563964377616387)\n\nWhat do you folks think about this comment? Do reply below", "704081": "it's not enough but it's not loosing game, it's something that should be done along side many other things.. cryptographic signatures are good but it needs the media platforms to use them to be useful.. if you go to a bad website and use that as your source (like a lot of people do) it won't make a difference because you decided to trust that source, it's about how much the receivers care about it more than how authentic it is imo", "705540": "I don't think it will be impossible for computers and I don't believe the two year timeline.", "705602": "I think he's right. As we learn how to beat a given level of deep fake technology, and the method of detection becomes known, the deepfake creation technology will evolve to fool it. Eventually, the fakes become, basically, perfect. I think that the end result will be that videos are no longer useable evidence - a pretty catastrophic endpoint.", "705603": "I agree with him. Its like a classic GAN situation, all this will do is create a better discriminator for the Deep Generative Models to beat. At the end, its at the source of data where authentication can be managed.", "706743": "Well, spotting \"an information bottleneck\" here is very theoretical. There was a time where classifying cats vs dogs was considered impossible using NP completeness arguments.", "707651": "Was classifying Cats vs Dogs so hard?", "709747": "Intellectuals have their opinions for sure, but the industy dominates.", "713921": "As long as humans can identify a deepfake by artifacts and such, we should definitely be able to identify by the current line of works. I think people will more follow Andrew's advice once it reaches that tipping point. In any case, this depends on the context of where we want to identify DeepFakes. Anonymous videos on Reddit arent gonna be verified by any means. So, AI will be useful there even if the videos can fool humans.", "729019": "How will *'real people'* claim that a digital object is real or fake? \n\nThere were two recent incidents where experts couldn't determine if the video was real or fake - one of them caused a [military coup (... a failed one though!)](https://www.motherjones.com/politics/2019/03/deepfake-gabon-ali-bongo) and the other a [minister to resign.](https://www.scmp.com/week-asia/politics/article/3014162/malaysian-minister-implicated-gay-sex-video-prompting) \n\nDigital signature for verifying/ authenticating a digital object is a great idea. But it's more from the copyright angle.\n\nDue to their far-reaching impact, DeepFakes are a social concern. And tools should be easily available to identify them. In fact we can equate it to the arms race in the Anti-Virus/Malware industry.", "1848481": "I agree with him. It is necessary to to sign content they created, It can embed some unique Identifier verified my profiles ids, only those IDs which created the digital objects, or upload the object, date ob digital object creation should embed in the data, It will be helpful to categories the resource as a new creation or old. At the end, its at the source of data where authentication can be managed." }, "source": "meta" }