I think it could be a good idea, but it might be quite tricky to implement : we are indeed working on differentiating images with small perturbations from their original ones, which is kind of exactly the same as adversarial examples (except the goal is not to trick a classifier). What do you think ?
", "messages": [ { "id": 890935, "postDate": "2020-06-17T19:18:20.240Z", "content": "I think it could be a good idea, but it might be quite tricky to implement : we are indeed working on differentiating images with small perturbations from their original ones, which is kind of exactly the same as adversarial examples (except the goal is not to trick a classifier). What do you think ?
", "rawMarkdown": "I think it could be a good idea, but it might be quite tricky to implement : we are indeed working on differentiating images with small perturbations from their original ones, which is kind of exactly the same as adversarial examples (except the goal is not to trick a classifier). What do you think ?", "votes": 3 }, { "id": 892880, "postDate": "2020-06-19T08:23:25.550Z", "content": "I have thought about using GANs for steganalysis but it seems most of them are only used for steganography. They usualy use this kind of architecture to hide the image and generate also an image from a distribution. However, the main disadvantage also is that there is no guarantee that minmax game will actually converge for GANs.
\n\nI think it is a lot of work but could be done for steganalysis. Good luck!\nHere is the paper: https://arxiv.org/pdf/1901.03892v2.pdf
", "rawMarkdown": "I have thought about using GANs for steganalysis but it seems most of them are only used for steganography. They usualy use this kind of architecture to hide the image and generate also an image from a distribution. However, the main disadvantage also is that there is no guarantee that minmax game will actually converge for GANs. \n\nI think it is a lot of work but could be done for steganalysis. Good luck!\nHere is the paper: https://arxiv.org/pdf/1901.03892v2.pdf", "votes": 1 }, { "id": 892108, "postDate": "2020-06-18T16:58:49.673Z", "content": "https://arxiv.org/abs/1904.12843\nThis suggests perturbations to be added to images at the direction of gradient. You will find a github link in it with implementation in tensorflow.\nLet me know if this helps.
", "rawMarkdown": "https://arxiv.org/abs/1904.12843\nThis suggests perturbations to be added to images at the direction of gradient. You will find a github link in it with implementation in tensorflow.\nLet me know if this helps. ", "votes": 1 }, { "id": 894071, "postDate": "2020-06-20T06:41:43.473Z", "content": "You are right. We need some new approaches to this competition. As you said it is quite tricky to implement, do you have any idea for detail? I want to give it a try.
", "rawMarkdown": "You are right. We need some new approaches to this competition. As you said it is quite tricky to implement, do you have any idea for detail? I want to give it a try.", "votes": 2 }, { "id": 891215, "postDate": "2020-06-18T02:52:58.423Z", "content": "are you sure, it is suitable for that? if yes, its new way to look GANs for me..
", "rawMarkdown": "are you sure, it is suitable for that? if yes, its new way to look GANs for me.." }, { "id": 895701, "postDate": "2020-06-21T14:48:54.267Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 892880, "author_name": "GuillemDelgado", "author_url": "", "post_date": "2020-06-19T08:23:25.550000", "content": "I have thought about using GANs for steganalysis but it seems most of them are only used for steganography. They usualy use this kind of architecture to hide the image and generate also an image from a distribution. However, the main disadvantage also is that there is no guarantee that minmax game will actually converge for GANs.
\n\nI think it is a lot of work but could be done for steganalysis. Good luck!\nHere is the paper: https://arxiv.org/pdf/1901.03892v2.pdf
", "votes": 1, "replies": [] }, { "id": 892108, "author_name": "Zaber Ibn Abdul Hakim", "author_url": "", "post_date": "2020-06-18T16:58:49.673000", "content": "https://arxiv.org/abs/1904.12843\nThis suggests perturbations to be added to images at the direction of gradient. You will find a github link in it with implementation in tensorflow.\nLet me know if this helps.
", "votes": 1, "replies": [] }, { "id": 894071, "author_name": "Johnny Lee", "author_url": "", "post_date": "2020-06-20T06:41:43.473000", "content": "You are right. We need some new approaches to this competition. As you said it is quite tricky to implement, do you have any idea for detail? I want to give it a try.
", "votes": 2, "replies": [] }, { "id": 891215, "author_name": "Uday Kumar Gurugubelli", "author_url": "", "post_date": "2020-06-18T02:52:58.423000", "content": "are you sure, it is suitable for that? if yes, its new way to look GANs for me..
", "votes": 0, "replies": [] }, { "id": 895701, "author_name": "", "author_url": "", "post_date": "2020-06-21T14:48:54.267000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "890935": "I think it could be a good idea, but it might be quite tricky to implement : we are indeed working on differentiating images with small perturbations from their original ones, which is kind of exactly the same as adversarial examples (except the goal is not to trick a classifier). What do you think ?", "892880": "I have thought about using GANs for steganalysis but it seems most of them are only used for steganography. They usualy use this kind of architecture to hide the image and generate also an image from a distribution. However, the main disadvantage also is that there is no guarantee that minmax game will actually converge for GANs. \n\nI think it is a lot of work but could be done for steganalysis. Good luck!\nHere is the paper: https://arxiv.org/pdf/1901.03892v2.pdf", "892108": "https://arxiv.org/abs/1904.12843\nThis suggests perturbations to be added to images at the direction of gradient. You will find a github link in it with implementation in tensorflow.\nLet me know if this helps. ", "894071": "You are right. We need some new approaches to this competition. As you said it is quite tricky to implement, do you have any idea for detail? I want to give it a try.", "891215": "are you sure, it is suitable for that? if yes, its new way to look GANs for me..", "895701": "" } }