![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2640743%2F67291870f62a5ac575c16f100a0a0327%2FSteg%20pap%20with%20code.png?generation=1588021925952421&alt=media\")
This competition will require us to detect how likely it is that an image contains secret information (a payload). It does not ask us to actually decode the data, but simply to express a level of confidence that it has a hidden message or not. In our submissions, higher or lower magnitudes indicate higher or lower confidence, while a positive or negative sign indicates that it it doesn't contain a hidden message (-) or it does (+). So 0.88 indicates it is fairly confident it contains a secret message, while -0.12 indicates it is not very confident that it does not contain a hidden message.
\n\nSince It is often difficult to find these messages unless you know to look for them, we are asked to find a better approach that can tell there is a hidden message based on patterns in the image itself. This is difficult because there are so many different ways to encode these secret messages in image data and so therefore there are many possible unique signatures of tampering. Moreover, there are many differences among ways to encode picture data itself (different cameras for example) which creates high variability in how the data is stored. Dealing with this level of variability has been challenging and so here we are, looking for innovative approaches to the problem.
\n\nRegarding submission evaluation: It is important to note that the evaluation for this contest will punish false positives more harshly and hence be careful of this when submitting. Here are some resources to help us all get started and learn about the field.
\n\n.Videos.
\n\nGreat introductory video on Steganalsysis - Really nice for understanding the basic concepts for those not in the field.\nhttps://youtu.be/HQGYpBZYtXs
\n\nComputerphile video on steganography\nhttps://youtu.be/TWEXCYQKyDc
\n\nDifficult to hear YouTube video on Steganalysis - very nice historical context and introduction:\nhttps://www.youtube.com/watch?v=rzTYWzHxLeY
\n\nSteganalysis by deep learning - 10 min YT video - Similar to our project.\nhttps://youtu.be/_o8nW3zF5r4
\n\nOlder pre-deep learning lecture that gives a sense of how many ways there are to encode messages. A ton! Wow!\nhttps://youtu.be/BQPkRlbVFEs\n.
\n\n** Great overview of statistical steganalysis ** - The process of detecting statistical traces of a secret communication.\nhttps://users.cs.fiu.edu/~fortega/spring17/df/research/Stenography%20II/Statistical%20Steganalysis.pdf
\n\nPapers with Code Search on steganalysis\nhttps://paperswithcode.com/search?q_meta=&q=steganalysis
\n\n
.
\n\nDeep Learning Hierarchical Representations for Image Steganalysis\nhttps://bv.univ-poitiers.fr/access/content/user/bdiall03/PhD_Image_Forensic_XLIM/Articles/CNN-MF/Deep%20Learning%20Steganalysis2017.pdf
\n\n.
\n\nLarge-scale JPEG image steganalysis using hybrid deep-learning framework\nhttps://arxiv.org/pdf/1611.03233.pdf](https://arxiv.org/pdf/1611.03233.pdf
\n\n\"Deep learning is a good steganalysis tool when embedding key is reused for different images, even if there is a cover source- mismatch\"
\n\nhttps://www.ingentaconnect.com/contentone/ist/ei/2016/00002016/00000008/art00014?crawler=true
\n\n.
\n\nArticles about the algorithms used to embed messages and evade detection:
\n\nJ-MiPod - one of the algorithms used to evade statistical detection.\n\"Steganography by Minimizing Statistical Detectability: The cases of JPEG and Color Images.\"\nhttps://hal-utt.archives-ouvertes.fr/hal-02542075/file/J_MiPOD_vPub.pdf
\n\nJUNIWARD - \n\"Steganalysis of Adaptive JPEG Steganography Using 2DGabor Filters\"\nhttps://dl.acm.org/doi/pdf/10.1145/2756601.2756608
\n\n.
\n\nJPEG
\n\nOn Improving Distortion Functions for JPEG Steganography\nhttps://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8554284
\n\n.
\n\nGANSs
\n\nAdversarial Embedding: A robust and elusive Steganography and Watermarking technique\nhttps://arxiv.org/pdf/1912.01487.pdf
\n\n.
\n\nHistorical overview:\n\"Deep Learning in steganog- 1 raphy and steganalysis from 2015 to 2018\"\nhttps://arxiv.org/pdf/1904.01444.pdf
\n\n.
\n\nRelevant Github Repositories:\n* https://github.com/RobinDavid/LSB-Steganography\n* https://github.com/harveyslash/Deep-Steganography\n* https://github.com/brijeshiitg/Pytorch-implementation-of-SRNet .Thanks @edosedgar !
\n\nPython Steganography packages\n* https://git.sr.ht/%7Ecedric/stegano
\n\nExternal Datasets (allowed in this competition so long as public)\n* https://github.com/YangzlTHU/IStego100K
\n\nFrom @edosedgar \n\"Deep Residual Network for Steganalysis of Digital Images\"\nhttp://www.ws.binghamton.edu/fridrich/Research/SRNet.pdf
\n\n.
\n\nFrom the previous incarnation of this competition - The Alaska1 Website:\nhttps://alaska.utt.fr/#acknowledgements\n.
\n\nFeel free to add research in the comments and I will quote you and add it above!
", "messages": [ { "id": 823746, "postDate": "2020-04-27T21:15:27.390Z", "content": "This competition will require us to detect how likely it is that an image contains secret information (a payload). It does not ask us to actually decode the data, but simply to express a level of confidence that it has a hidden message or not. In our submissions, higher or lower magnitudes indicate higher or lower confidence, while a positive or negative sign indicates that it it doesn't contain a hidden message (-) or it does (+). So 0.88 indicates it is fairly confident it contains a secret message, while -0.12 indicates it is not very confident that it does not contain a hidden message.
\n\nSince It is often difficult to find these messages unless you know to look for them, we are asked to find a better approach that can tell there is a hidden message based on patterns in the image itself. This is difficult because there are so many different ways to encode these secret messages in image data and so therefore there are many possible unique signatures of tampering. Moreover, there are many differences among ways to encode picture data itself (different cameras for example) which creates high variability in how the data is stored. Dealing with this level of variability has been challenging and so here we are, looking for innovative approaches to the problem.
\n\nRegarding submission evaluation: It is important to note that the evaluation for this contest will punish false positives more harshly and hence be careful of this when submitting. Here are some resources to help us all get started and learn about the field.
\n\n.Videos.
\n\nGreat introductory video on Steganalsysis - Really nice for understanding the basic concepts for those not in the field.\nhttps://youtu.be/HQGYpBZYtXs
\n\nComputerphile video on steganography\nhttps://youtu.be/TWEXCYQKyDc
\n\nDifficult to hear YouTube video on Steganalysis - very nice historical context and introduction:\nhttps://www.youtube.com/watch?v=rzTYWzHxLeY
\n\nSteganalysis by deep learning - 10 min YT video - Similar to our project.\nhttps://youtu.be/_o8nW3zF5r4
\n\nOlder pre-deep learning lecture that gives a sense of how many ways there are to encode messages. A ton! Wow!\nhttps://youtu.be/BQPkRlbVFEs\n.
\n\n** Great overview of statistical steganalysis ** - The process of detecting statistical traces of a secret communication.\nhttps://users.cs.fiu.edu/~fortega/spring17/df/research/Stenography%20II/Statistical%20Steganalysis.pdf
\n\nPapers with Code Search on steganalysis\nhttps://paperswithcode.com/search?q_meta=&q=steganalysis
\n\n
.
\n\nDeep Learning Hierarchical Representations for Image Steganalysis\nhttps://bv.univ-poitiers.fr/access/content/user/bdiall03/PhD_Image_Forensic_XLIM/Articles/CNN-MF/Deep%20Learning%20Steganalysis2017.pdf
\n\n.
\n\nLarge-scale JPEG image steganalysis using hybrid deep-learning framework\nhttps://arxiv.org/pdf/1611.03233.pdf](https://arxiv.org/pdf/1611.03233.pdf
\n\n\"Deep learning is a good steganalysis tool when embedding key is reused for different images, even if there is a cover source- mismatch\"
\n\nhttps://www.ingentaconnect.com/contentone/ist/ei/2016/00002016/00000008/art00014?crawler=true
\n\n.
\n\nArticles about the algorithms used to embed messages and evade detection:
\n\nJ-MiPod - one of the algorithms used to evade statistical detection.\n\"Steganography by Minimizing Statistical Detectability: The cases of JPEG and Color Images.\"\nhttps://hal-utt.archives-ouvertes.fr/hal-02542075/file/J_MiPOD_vPub.pdf
\n\nJUNIWARD - \n\"Steganalysis of Adaptive JPEG Steganography Using 2DGabor Filters\"\nhttps://dl.acm.org/doi/pdf/10.1145/2756601.2756608
\n\n.
\n\nJPEG
\n\nOn Improving Distortion Functions for JPEG Steganography\nhttps://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8554284
\n\n.
\n\nGANSs
\n\nAdversarial Embedding: A robust and elusive Steganography and Watermarking technique\nhttps://arxiv.org/pdf/1912.01487.pdf
\n\n.
\n\nHistorical overview:\n\"Deep Learning in steganog- 1 raphy and steganalysis from 2015 to 2018\"\nhttps://arxiv.org/pdf/1904.01444.pdf
\n\n.
\n\nRelevant Github Repositories:\n* https://github.com/RobinDavid/LSB-Steganography\n* https://github.com/harveyslash/Deep-Steganography\n* https://github.com/brijeshiitg/Pytorch-implementation-of-SRNet .Thanks @edosedgar !
\n\nPython Steganography packages\n* https://git.sr.ht/%7Ecedric/stegano
\n\nExternal Datasets (allowed in this competition so long as public)\n* https://github.com/YangzlTHU/IStego100K
\n\nFrom @edosedgar \n\"Deep Residual Network for Steganalysis of Digital Images\"\nhttp://www.ws.binghamton.edu/fridrich/Research/SRNet.pdf
\n\n.
\n\nFrom the previous incarnation of this competition - The Alaska1 Website:\nhttps://alaska.utt.fr/#acknowledgements\n.
\n\nFeel free to add research in the comments and I will quote you and add it above!
", "rawMarkdown": "This competition will require us to **detect how likely it is that an image contains secret information (a payload).** It does not ask us to actually decode the data, but simply to express a level of confidence that it has a hidden message or not. In our submissions, **higher or lower magnitudes indicate higher or lower confidence**, while **a positive or negative sign indicates that it it doesn't contain a hidden message (-) or it does (+)**. So 0.88 indicates it is fairly confident it contains a secret message, while -0.12 indicates it is not very confident that it does not contain a hidden message.\n\nSince It is often difficult to find these messages unless you know to look for them, we are asked to find a better approach that can tell there is a hidden message based on patterns in the image itself. This is difficult because there are so many different ways to encode these secret messages in image data and so therefore **there are many possible unique signatures of tampering**. Moreover, there are many differences among ways to encode picture data itself (different cameras for example) which creates **high variability in how the data is stored**. Dealing with this level of variability has been challenging and so here we are, looking for innovative approaches to the problem.\n\n**Regarding submission evaluation:** It is important to note that the evaluation for this contest will **punish false positives** more harshly and hence be careful of this when submitting. Here are some resources to help us all get started and learn about the field.\n\n.Videos.\n\n**Great introductory video on Steganalsysis** - Really nice for understanding the basic concepts for those not in the field.\nhttps://youtu.be/HQGYpBZYtXs\n\nComputerphile video on steganography\nhttps://youtu.be/TWEXCYQKyDc\n\nDifficult to hear YouTube video on Steganalysis - very nice historical context and introduction:\nhttps://www.youtube.com/watch?v=rzTYWzHxLeY\n\nSteganalysis by deep learning - 10 min YT video - Similar to our project.\nhttps://youtu.be/_o8nW3zF5r4\n\nOlder pre-deep learning lecture that gives a sense of how many ways there are to encode messages. A ton! Wow!\nhttps://youtu.be/BQPkRlbVFEs\n.\n\n\n** Great overview of statistical steganalysis ** - The process of detecting statistical traces of a secret communication.\nhttps://users.cs.fiu.edu/~fortega/spring17/df/research/Stenography%20II/Statistical%20Steganalysis.pdf\n\n**Papers with Code Search on steganalysis**\n[https://paperswithcode.com/search?q_meta=&q=steganalysis](https://paperswithcode.com/search?q_meta=&q=steganalysis)\n\n\n[](https://paperswithcode.com/search?q_meta=&q=steganalysis)\n\n.\n\nDeep Learning Hierarchical Representations for Image Steganalysis\nhttps://bv.univ-poitiers.fr/access/content/user/bdiall03/PhD_Image_Forensic_XLIM/Articles/CNN-MF/Deep%20Learning%20Steganalysis2017.pdf\n\n.\n\nLarge-scale JPEG image steganalysis using hybrid deep-learning framework\nhttps://arxiv.org/pdf/1611.03233.pdf](https://arxiv.org/pdf/1611.03233.pdf\n\n\"Deep learning is a good steganalysis tool when embedding key is reused for different images, even if there is a cover source- mismatch\"\n\nhttps://www.ingentaconnect.com/contentone/ist/ei/2016/00002016/00000008/art00014?crawler=true\n\n.\n\n**Articles about the algorithms used to embed messages and evade detection:**\n\nJ-MiPod - one of the algorithms used to evade statistical detection.\n\"Steganography by Minimizing Statistical Detectability: The cases of JPEG and Color Images.\"\nhttps://hal-utt.archives-ouvertes.fr/hal-02542075/file/J_MiPOD_vPub.pdf\n\nJUNIWARD - \n\"Steganalysis of Adaptive JPEG Steganography Using 2DGabor Filters\"\nhttps://dl.acm.org/doi/pdf/10.1145/2756601.2756608\n\n.\n\nJPEG\n\nOn Improving Distortion Functions for JPEG Steganography\nhttps://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8554284\n\n.\n\nGANSs\n\nAdversarial Embedding: A robust and elusive Steganography and Watermarking technique\nhttps://arxiv.org/pdf/1912.01487.pdf\n\n.\n\nHistorical overview:\n\"Deep Learning in steganog- 1 raphy and steganalysis from 2015 to 2018\"\nhttps://arxiv.org/pdf/1904.01444.pdf\n\n.\n\nRelevant Github Repositories:\n* https://github.com/RobinDavid/LSB-Steganography\n* https://github.com/harveyslash/Deep-Steganography\n* https://github.com/brijeshiitg/Pytorch-implementation-of-SRNet .Thanks @edosedgar ! \n\nPython Steganography packages\n* https://git.sr.ht/%7Ecedric/stegano\n\nExternal Datasets (allowed in this competition so long as public)\n* https://github.com/YangzlTHU/IStego100K\n\n\nFrom @edosedgar \n\"Deep Residual Network for Steganalysis of Digital Images\"\nhttp://www.ws.binghamton.edu/fridrich/Research/SRNet.pdf\n\n.\n\nFrom the previous incarnation of this competition - The Alaska1 Website:\nhttps://alaska.utt.fr/#acknowledgements\n.\n\n\nFeel free to add research in the comments and I will quote you and add it above!", "votes": 59 }, { "id": 824148, "postDate": "2020-04-28T07:31:04.440Z", "content": "Wow ! Thank you so much for this terrific post !\nI am frankly amazed (especially since you wrote it within an hour !!) ; many good source, all information are 100% accurate and relevant. Some references I actually did not know and I will reuse for my students 😄
\n\nI would not have had any better !\nTwo minor information perhaps, one obvious, beware that image steganography has been mostly developed by academic community for spatial domain images (uncompressed / pixels) ; here we use JPEG images to be way more realistic\nSecond, the third embedding method used in UERD \"Using statistical image model for JPEG steganography: uniform embedding revisited\" , unfortunately the paper is not publicly available (I admit I did not pay attention to this, but I am sure you all know a website to find such papers ... 😄 )\nI can provide embedding source code, but those are written in Matlab .* file 😢
\n\nThanks a lot again tpmeli !\nreally looking forward if you join the competition and how you will perform.
", "rawMarkdown": "Wow ! Thank you so much for this terrific post !\nI am frankly amazed (especially since you wrote it within an hour !!) ; many good source, all information are 100% accurate and relevant. Some references I actually did not know and I will reuse for my students 😄 \n\nI would not have had any better !\nTwo minor information perhaps, one obvious, beware that image steganography has been mostly developed by academic community for spatial domain images (uncompressed / pixels) ; here we use JPEG images to be way more realistic\nSecond, the third embedding method used in UERD \"Using statistical image model for JPEG steganography: uniform embedding revisited\" , unfortunately the paper is not publicly available (I admit I did not pay attention to this, but I am sure you all know a website to find such papers ... 😄 )\nI can provide embedding source code, but those are written in Matlab .* file 😢 \n\nThanks a lot again tpmeli !\nreally looking forward if you join the competition and how you will perform.", "votes": 3, "replies": [ { "id": 824225, "postDate": "2020-04-28T08:34:54.887Z", "content": "Thanks!
\n\nIt would be great to have the programs used to hide the information. Even if they are written in Matlab ...
", "rawMarkdown": "Thanks!\n\nIt would be great to have the programs used to hide the information. Even if they are written in Matlab ..." }, { "id": 824376, "postDate": "2020-04-28T10:27:16.393Z", "rawMarkdown": "", "isDeleted": true }, { "id": 824728, "postDate": "2020-04-28T15:01:23.447Z", "content": "Hi Remi, thanks so much! I give credit to the awesome Kagglers whose example I am following and learning from. :). Thank you for your insights about JPEG compression and the other paper. Cheers and thanks for your encouragement!
", "rawMarkdown": "Hi Remi, thanks so much! I give credit to the awesome Kagglers whose example I am following and learning from. :). Thank you for your insights about JPEG compression and the other paper. Cheers and thanks for your encouragement!", "votes": 1 } ] }, { "id": 835140, "postDate": "2020-05-06T04:52:09.597Z", "content": "great
", "rawMarkdown": "great", "votes": 1 }, { "id": 823912, "postDate": "2020-04-28T02:31:00.527Z", "content": "Thanks for all the resources. Upvoted!
", "rawMarkdown": "Thanks for all the resources. Upvoted!", "votes": 1, "replies": [ { "id": 823915, "postDate": "2020-04-28T02:41:53.947Z", "content": "Thanks Ark!
", "rawMarkdown": "Thanks Ark!", "votes": 1 } ] }, { "id": 837019, "postDate": "2020-05-07T12:58:19.723Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 837050, "postDate": "2020-05-07T13:35:00.430Z", "content": "If you are unsure, you can check on the rules page for this competition. It states the following:\n \"...External Data. You may use data other than the Competition Data (“External Data”) to develop and test your models and Submissions. However, you will (i) ensure the External Data is available to use by all participants of the competition for purposes of the competition at no cost to the other participants and (ii) post such access to the External Data for the participants to the official competition forum prior to the Entry Deadline.\"\n-https://www.kaggle.com/c/alaska2-image-steganalysis/rules
", "rawMarkdown": "If you are unsure, you can check on the rules page for this competition. It states the following:\n \"...External Data. You may use data other than the Competition Data (“External Data”) to develop and test your models and Submissions. However, you will (i) ensure the External Data is available to use by all participants of the competition for purposes of the competition at no cost to the other participants and (ii) post such access to the External Data for the participants to the official competition forum prior to the Entry Deadline.\"\n-https://www.kaggle.com/c/alaska2-image-steganalysis/rules", "votes": 1 }, { "id": 837089, "postDate": "2020-05-07T14:06:22.483Z", "rawMarkdown": "", "isDeleted": true }, { "id": 837276, "postDate": "2020-05-07T16:51:18.007Z", "content": "Yes, I saw that too. I think this was referring to the difficulty of Remi sharing a particular dataset with us, not saying that it was against the rules to use external data sources. It's kind of like saying: \"sure we could give them to you, but you probably don't need them.\" This may indeed be correct, but it's hard to tell in advance since transfer learning often provides non-intuitive results.
", "rawMarkdown": "Yes, I saw that too. I think this was referring to the difficulty of Remi sharing a particular dataset with us, not saying that it was against the rules to use external data sources. It's kind of like saying: \"sure we could give them to you, but you probably don't need them.\" This may indeed be correct, but it's hard to tell in advance since transfer learning often provides non-intuitive results.", "votes": 2 } ] }, { "id": 824395, "postDate": "2020-04-28T10:41:05.347Z", "rawMarkdown": "", "votes": 1, "isDeleted": true, "replies": [ { "id": 824726, "postDate": "2020-04-28T15:00:08.473Z", "content": "This is awesome, thank you!
", "rawMarkdown": "This is awesome, thank you!", "votes": 1 } ] }, { "id": 823766, "postDate": "2020-04-27T21:47:28.717Z", "rawMarkdown": "", "isDeleted": true }, { "id": 823760, "postDate": "2020-04-27T21:42:13.663Z", "rawMarkdown": "", "votes": 2, "isDeleted": true, "replies": [ { "id": 823765, "postDate": "2020-04-27T21:46:26.113Z", "content": "Thanks, I posted it. :)
", "rawMarkdown": "Thanks, I posted it. :) ", "votes": 1 } ] }, { "id": 870882, "postDate": "2020-06-02T02:26:31.747Z", "content": "Thanks for the resources!!
", "rawMarkdown": "Thanks for the resources!!" } ], "comments": [ { "id": 824148, "author_name": "Rémi Cogranne", "author_url": "", "post_date": "2020-04-28T07:31:04.440000", "content": "Wow ! Thank you so much for this terrific post !\nI am frankly amazed (especially since you wrote it within an hour !!) ; many good source, all information are 100% accurate and relevant. Some references I actually did not know and I will reuse for my students 😄
\n\nI would not have had any better !\nTwo minor information perhaps, one obvious, beware that image steganography has been mostly developed by academic community for spatial domain images (uncompressed / pixels) ; here we use JPEG images to be way more realistic\nSecond, the third embedding method used in UERD \"Using statistical image model for JPEG steganography: uniform embedding revisited\" , unfortunately the paper is not publicly available (I admit I did not pay attention to this, but I am sure you all know a website to find such papers ... 😄 )\nI can provide embedding source code, but those are written in Matlab .* file 😢
\n\nThanks a lot again tpmeli !\nreally looking forward if you join the competition and how you will perform.
", "votes": 3, "replies": [ { "id": 824225, "author_name": "Daniel Lerch", "author_url": "", "post_date": "2020-04-28T08:34:54.887000", "content": "Thanks!
\n\nIt would be great to have the programs used to hide the information. Even if they are written in Matlab ...
", "votes": 0, "replies": [] }, { "id": 824376, "author_name": "", "author_url": "", "post_date": "2020-04-28T10:27:16.393000", "content": "", "votes": 0, "replies": [] }, { "id": 824728, "author_name": "Tom M", "author_url": "", "post_date": "2020-04-28T15:01:23.447000", "content": "Hi Remi, thanks so much! I give credit to the awesome Kagglers whose example I am following and learning from. :). Thank you for your insights about JPEG compression and the other paper. Cheers and thanks for your encouragement!
", "votes": 1, "replies": [] } ] }, { "id": 835140, "author_name": "temptmp", "author_url": "", "post_date": "2020-05-06T04:52:09.597000", "content": "great
", "votes": 1, "replies": [] }, { "id": 823912, "author_name": "ark", "author_url": "", "post_date": "2020-04-28T02:31:00.527000", "content": "Thanks for all the resources. Upvoted!
", "votes": 1, "replies": [ { "id": 823915, "author_name": "Tom M", "author_url": "", "post_date": "2020-04-28T02:41:53.947000", "content": "Thanks Ark!
", "votes": 1, "replies": [] } ] }, { "id": 837019, "author_name": "", "author_url": "", "post_date": "2020-05-07T12:58:19.723000", "content": "", "votes": 0, "replies": [ { "id": 837050, "author_name": "Tom M", "author_url": "", "post_date": "2020-05-07T13:35:00.430000", "content": "If you are unsure, you can check on the rules page for this competition. It states the following:\n \"...External Data. You may use data other than the Competition Data (“External Data”) to develop and test your models and Submissions. However, you will (i) ensure the External Data is available to use by all participants of the competition for purposes of the competition at no cost to the other participants and (ii) post such access to the External Data for the participants to the official competition forum prior to the Entry Deadline.\"\n-https://www.kaggle.com/c/alaska2-image-steganalysis/rules
", "votes": 1, "replies": [] }, { "id": 837089, "author_name": "", "author_url": "", "post_date": "2020-05-07T14:06:22.483000", "content": "", "votes": 0, "replies": [] }, { "id": 837276, "author_name": "Tom M", "author_url": "", "post_date": "2020-05-07T16:51:18.007000", "content": "Yes, I saw that too. I think this was referring to the difficulty of Remi sharing a particular dataset with us, not saying that it was against the rules to use external data sources. It's kind of like saying: \"sure we could give them to you, but you probably don't need them.\" This may indeed be correct, but it's hard to tell in advance since transfer learning often provides non-intuitive results.
", "votes": 2, "replies": [] } ] }, { "id": 824395, "author_name": "", "author_url": "", "post_date": "2020-04-28T10:41:05.347000", "content": "", "votes": 1, "replies": [ { "id": 824726, "author_name": "Tom M", "author_url": "", "post_date": "2020-04-28T15:00:08.473000", "content": "This is awesome, thank you!
", "votes": 1, "replies": [] } ] }, { "id": 823766, "author_name": "", "author_url": "", "post_date": "2020-04-27T21:47:28.717000", "content": "", "votes": 0, "replies": [] }, { "id": 823760, "author_name": "", "author_url": "", "post_date": "2020-04-27T21:42:13.663000", "content": "", "votes": 2, "replies": [ { "id": 823765, "author_name": "Tom M", "author_url": "", "post_date": "2020-04-27T21:46:26.113000", "content": "Thanks, I posted it. :)
", "votes": 1, "replies": [] } ] }, { "id": 870882, "author_name": "Brown Fox", "author_url": "", "post_date": "2020-06-02T02:26:31.747000", "content": "Thanks for the resources!!
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "823746": "This competition will require us to **detect how likely it is that an image contains secret information (a payload).** It does not ask us to actually decode the data, but simply to express a level of confidence that it has a hidden message or not. In our submissions, **higher or lower magnitudes indicate higher or lower confidence**, while **a positive or negative sign indicates that it it doesn't contain a hidden message (-) or it does (+)**. So 0.88 indicates it is fairly confident it contains a secret message, while -0.12 indicates it is not very confident that it does not contain a hidden message.\n\nSince It is often difficult to find these messages unless you know to look for them, we are asked to find a better approach that can tell there is a hidden message based on patterns in the image itself. This is difficult because there are so many different ways to encode these secret messages in image data and so therefore **there are many possible unique signatures of tampering**. Moreover, there are many differences among ways to encode picture data itself (different cameras for example) which creates **high variability in how the data is stored**. Dealing with this level of variability has been challenging and so here we are, looking for innovative approaches to the problem.\n\n**Regarding submission evaluation:** It is important to note that the evaluation for this contest will **punish false positives** more harshly and hence be careful of this when submitting. Here are some resources to help us all get started and learn about the field.\n\n.Videos.\n\n**Great introductory video on Steganalsysis** - Really nice for understanding the basic concepts for those not in the field.\nhttps://youtu.be/HQGYpBZYtXs\n\nComputerphile video on steganography\nhttps://youtu.be/TWEXCYQKyDc\n\nDifficult to hear YouTube video on Steganalysis - very nice historical context and introduction:\nhttps://www.youtube.com/watch?v=rzTYWzHxLeY\n\nSteganalysis by deep learning - 10 min YT video - Similar to our project.\nhttps://youtu.be/_o8nW3zF5r4\n\nOlder pre-deep learning lecture that gives a sense of how many ways there are to encode messages. A ton! Wow!\nhttps://youtu.be/BQPkRlbVFEs\n.\n\n\n** Great overview of statistical steganalysis ** - The process of detecting statistical traces of a secret communication.\nhttps://users.cs.fiu.edu/~fortega/spring17/df/research/Stenography%20II/Statistical%20Steganalysis.pdf\n\n**Papers with Code Search on steganalysis**\n[https://paperswithcode.com/search?q_meta=&q=steganalysis](https://paperswithcode.com/search?q_meta=&q=steganalysis)\n\n\n[](https://paperswithcode.com/search?q_meta=&q=steganalysis)\n\n.\n\nDeep Learning Hierarchical Representations for Image Steganalysis\nhttps://bv.univ-poitiers.fr/access/content/user/bdiall03/PhD_Image_Forensic_XLIM/Articles/CNN-MF/Deep%20Learning%20Steganalysis2017.pdf\n\n.\n\nLarge-scale JPEG image steganalysis using hybrid deep-learning framework\nhttps://arxiv.org/pdf/1611.03233.pdf](https://arxiv.org/pdf/1611.03233.pdf\n\n\"Deep learning is a good steganalysis tool when embedding key is reused for different images, even if there is a cover source- mismatch\"\n\nhttps://www.ingentaconnect.com/contentone/ist/ei/2016/00002016/00000008/art00014?crawler=true\n\n.\n\n**Articles about the algorithms used to embed messages and evade detection:**\n\nJ-MiPod - one of the algorithms used to evade statistical detection.\n\"Steganography by Minimizing Statistical Detectability: The cases of JPEG and Color Images.\"\nhttps://hal-utt.archives-ouvertes.fr/hal-02542075/file/J_MiPOD_vPub.pdf\n\nJUNIWARD - \n\"Steganalysis of Adaptive JPEG Steganography Using 2DGabor Filters\"\nhttps://dl.acm.org/doi/pdf/10.1145/2756601.2756608\n\n.\n\nJPEG\n\nOn Improving Distortion Functions for JPEG Steganography\nhttps://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8554284\n\n.\n\nGANSs\n\nAdversarial Embedding: A robust and elusive Steganography and Watermarking technique\nhttps://arxiv.org/pdf/1912.01487.pdf\n\n.\n\nHistorical overview:\n\"Deep Learning in steganog- 1 raphy and steganalysis from 2015 to 2018\"\nhttps://arxiv.org/pdf/1904.01444.pdf\n\n.\n\nRelevant Github Repositories:\n* https://github.com/RobinDavid/LSB-Steganography\n* https://github.com/harveyslash/Deep-Steganography\n* https://github.com/brijeshiitg/Pytorch-implementation-of-SRNet .Thanks @edosedgar ! \n\nPython Steganography packages\n* https://git.sr.ht/%7Ecedric/stegano\n\nExternal Datasets (allowed in this competition so long as public)\n* https://github.com/YangzlTHU/IStego100K\n\n\nFrom @edosedgar \n\"Deep Residual Network for Steganalysis of Digital Images\"\nhttp://www.ws.binghamton.edu/fridrich/Research/SRNet.pdf\n\n.\n\nFrom the previous incarnation of this competition - The Alaska1 Website:\nhttps://alaska.utt.fr/#acknowledgements\n.\n\n\nFeel free to add research in the comments and I will quote you and add it above!", "824148": "Wow ! Thank you so much for this terrific post !\nI am frankly amazed (especially since you wrote it within an hour !!) ; many good source, all information are 100% accurate and relevant. Some references I actually did not know and I will reuse for my students 😄 \n\nI would not have had any better !\nTwo minor information perhaps, one obvious, beware that image steganography has been mostly developed by academic community for spatial domain images (uncompressed / pixels) ; here we use JPEG images to be way more realistic\nSecond, the third embedding method used in UERD \"Using statistical image model for JPEG steganography: uniform embedding revisited\" , unfortunately the paper is not publicly available (I admit I did not pay attention to this, but I am sure you all know a website to find such papers ... 😄 )\nI can provide embedding source code, but those are written in Matlab .* file 😢 \n\nThanks a lot again tpmeli !\nreally looking forward if you join the competition and how you will perform.", "835140": "great", "823912": "Thanks for all the resources. Upvoted!", "837019": "", "824395": "", "823766": "", "823760": "", "870882": "Thanks for the resources!!" } }