Dr. @remicogranne,
\n\nThank you for organizing this competition. A week ago, I knew nothing of this field. Not much has changed yet 😬😅😠but after having read so many papers and watched so many videos, I now am at least aware of its existence and have started familiarizing myself with some of the jargon.
\n\nIf possible, could you share a bit about what will happen after the competition ends? For example, what will the winning entries of this competition be used for exactly? The WIFS website is pretty vague and essentially purports that the goal here is simply to introduce / involve the greater ML community to steganography and steganalysis, and to expand the boundaries of state-of-the-art.
\n\nI stem from the applied DS side (I'm a contractor, not a researcher!) so whenever I see money on the table, there's usually some party pushing an agenda or trying to achieve some measurable objective. I don't know how ALASKA2 was funded, and it's not like I'm entitled to that information. But as I discover more about steganalysis specifically, I'm also curious how the results of this effort will be applied, who the groups organizing this effort are, and what they're all about. Since you are very deep in that world, any additional information you can share on that front would be greatly appreciated.
\n\nThank you!
", "messages": [ { "id": "913807", "postDate": "07/03/2020 12:51:58", "content": "Dr. @remicogranne,
\n\nThank you for organizing this competition. A week ago, I knew nothing of this field. Not much has changed yet 😬😅😠but after having read so many papers and watched so many videos, I now am at least aware of its existence and have started familiarizing myself with some of the jargon.
\n\nIf possible, could you share a bit about what will happen after the competition ends? For example, what will the winning entries of this competition be used for exactly? The WIFS website is pretty vague and essentially purports that the goal here is simply to introduce / involve the greater ML community to steganography and steganalysis, and to expand the boundaries of state-of-the-art.
\n\nI stem from the applied DS side (I'm a contractor, not a researcher!) so whenever I see money on the table, there's usually some party pushing an agenda or trying to achieve some measurable objective. I don't know how ALASKA2 was funded, and it's not like I'm entitled to that information. But as I discover more about steganalysis specifically, I'm also curious how the results of this effort will be applied, who the groups organizing this effort are, and what they're all about. Since you are very deep in that world, any additional information you can share on that front would be greatly appreciated.
\n\nThank you!
", "rawMarkdown": "Dr. @remicogranne,\n\nThank you for organizing this competition. A week ago, I knew nothing of this field. Not much has changed yet 😬😅😠but after having read so many papers and watched so many videos, I now am at least aware of its existence and have started familiarizing myself with some of the jargon.\n\nIf possible, could you share a bit about what will happen after the competition ends? For example, what will the winning entries of this competition be used for exactly? The [WIFS website](https://www.wifs2020.nyu.edu/home) is pretty vague and essentially purports that the goal here is simply to introduce / involve the greater ML community to steganography and steganalysis, and to expand the boundaries of state-of-the-art.\n\nI stem from the applied DS side (I'm a contractor, not a researcher!) so whenever I see money on the table, there's usually some party pushing an agenda or trying to achieve some measurable objective. I don't know how ALASKA2 was funded, and it's not like I'm entitled to that information. But as I discover more about steganalysis specifically, I'm also curious how the results of this effort will be applied, who the groups organizing this effort are, and what they're all about. Since you are very deep in that world, any additional information you can share on that front would be greatly appreciated.\n\nThank you!", "votes": null }, { "id": "914226", "postDate": "07/03/2020 17:26:45", "content": "This is a question for Remi but I will take a stab. My understanding is that the main goal of this competition is to push the state of the art in steganalysis in more real-life conditions, e.g., for current best stego schemes that are not necessarily available to the steganalyst, variable and unknown payload size, and when the embedding is spread among color channels. The detectors we build (will have built) are unlikely to be used in practice IMHO as only three embedding schemes are involved. There are dozens of steganographic schemes, many old ones, that are in fact much more likely to be used in practice by steganography users, such as F5, OutGuess, JP Hide&Seek, Steghide, Jsteg, etc, etc. Plus, this competition does not address the problem of how to steganalyze larger images. What if you have 20 megapixel image, for example? How would you make a decision about it using your detector? You should not resize as that would be highly suboptimal, so you may apply your detector to individual tiles of the image and \"fuse\" the decisions in some manner. This important aspect is completely missing here. Moreover, we only covered three quality factors. There are JPEGs with custom quantization tables, JPEGs with different chroma subsampling, and different encoding (RGB vs. YCrCb), etc. For a detector that could be used in real life, all this (immense) complexity would have to addressed.
\n\nHaving said all this, this competition has already brought a lot of fresh wind to the field of steganography that will lead to many new publications when properly studied in a controlled environment. Remember, academicians want understanding, they are far less interested in building actual detectors. So, to everyone competing here : \"great job, guys, you are helping our field to flourish by trying ideas that stego experts may not venture into.\" Apparently, engineering detectors is a job best left to deep learning experts who know just a little about steganography rather than stego experts who know a little about deep learning. This, by itself, is useful to know. Technology keeps on evolving, we retool, relearn, and embrace new paradigms, to ask more questions as we strive to understand. The better detectors will reshape the state of the art in steganography as yesterday's kings will fall and new kings emerge. The question of the best steganographic scheme will be revisited with this new generation of detectors. So far, it looks like the 7-year old J-UNIWARD may just survive the onslaught the best while the new kid on the block, a better theoretically founded J-MiPOD, may need to be retuned to be more secure in this never-ending game of cat and mouse.
\n\nOnce more, thanks everyone for participating. It is much appreciated whoever wins by a hair width in the end.
", "rawMarkdown": "This is a question for Remi but I will take a stab. My understanding is that the main goal of this competition is to push the state of the art in steganalysis in more real-life conditions, e.g., for current best stego schemes that are not necessarily available to the steganalyst, variable and unknown payload size, and when the embedding is spread among color channels. The detectors we build (will have built) are unlikely to be used in practice IMHO as only three embedding schemes are involved. There are dozens of steganographic schemes, many old ones, that are in fact much more likely to be used in practice by steganography users, such as F5, OutGuess, JP Hide&Seek, Steghide, Jsteg, etc, etc. Plus, this competition does not address the problem of how to steganalyze larger images. What if you have 20 megapixel image, for example? How would you make a decision about it using your detector? You should not resize as that would be highly suboptimal, so you may apply your detector to individual tiles of the image and \"fuse\" the decisions in some manner. This important aspect is completely missing here. Moreover, we only covered three quality factors. There are JPEGs with custom quantization tables, JPEGs with different chroma subsampling, and different encoding (RGB vs. YCrCb), etc. For a detector that could be used in real life, all this (immense) complexity would have to addressed.\n\nHaving said all this, this competition has already brought a lot of fresh wind to the field of steganography that will lead to many new publications when properly studied in a controlled environment. Remember, academicians want understanding, they are far less interested in building actual detectors. So, to everyone competing here : \"great job, guys, you are helping our field to flourish by trying ideas that stego experts may not venture into.\" Apparently, engineering detectors is a job best left to deep learning experts who know just a little about steganography rather than stego experts who know a little about deep learning. This, by itself, is useful to know. Technology keeps on evolving, we retool, relearn, and embrace new paradigms, to ask more questions as we strive to understand. The better detectors will reshape the state of the art in steganography as yesterday's kings will fall and new kings emerge. The question of the best steganographic scheme will be revisited with this new generation of detectors. So far, it looks like the 7-year old J-UNIWARD may just survive the onslaught the best while the new kid on the block, a better theoretically founded J-MiPOD, may need to be retuned to be more secure in this never-ending game of cat and mouse. \n\nOnce more, thanks everyone for participating. It is much appreciated whoever wins by a hair width in the end.", "votes": null } ], "comments": [ { "id": 914226, "author_name": "agnethajf", "author_url": "", "post_date": "07/03/2020 17:26:45", "content": "This is a question for Remi but I will take a stab. My understanding is that the main goal of this competition is to push the state of the art in steganalysis in more real-life conditions, e.g., for current best stego schemes that are not necessarily available to the steganalyst, variable and unknown payload size, and when the embedding is spread among color channels. The detectors we build (will have built) are unlikely to be used in practice IMHO as only three embedding schemes are involved. There are dozens of steganographic schemes, many old ones, that are in fact much more likely to be used in practice by steganography users, such as F5, OutGuess, JP Hide&Seek, Steghide, Jsteg, etc, etc. Plus, this competition does not address the problem of how to steganalyze larger images. What if you have 20 megapixel image, for example? How would you make a decision about it using your detector? You should not resize as that would be highly suboptimal, so you may apply your detector to individual tiles of the image and \"fuse\" the decisions in some manner. This important aspect is completely missing here. Moreover, we only covered three quality factors. There are JPEGs with custom quantization tables, JPEGs with different chroma subsampling, and different encoding (RGB vs. YCrCb), etc. For a detector that could be used in real life, all this (immense) complexity would have to addressed.
\n\nHaving said all this, this competition has already brought a lot of fresh wind to the field of steganography that will lead to many new publications when properly studied in a controlled environment. Remember, academicians want understanding, they are far less interested in building actual detectors. So, to everyone competing here : \"great job, guys, you are helping our field to flourish by trying ideas that stego experts may not venture into.\" Apparently, engineering detectors is a job best left to deep learning experts who know just a little about steganography rather than stego experts who know a little about deep learning. This, by itself, is useful to know. Technology keeps on evolving, we retool, relearn, and embrace new paradigms, to ask more questions as we strive to understand. The better detectors will reshape the state of the art in steganography as yesterday's kings will fall and new kings emerge. The question of the best steganographic scheme will be revisited with this new generation of detectors. So far, it looks like the 7-year old J-UNIWARD may just survive the onslaught the best while the new kid on the block, a better theoretically founded J-MiPOD, may need to be retuned to be more secure in this never-ending game of cat and mouse.
\n\nOnce more, thanks everyone for participating. It is much appreciated whoever wins by a hair width in the end.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "913807": "Dr. @remicogranne,\n\nThank you for organizing this competition. A week ago, I knew nothing of this field. Not much has changed yet 😬😅😠but after having read so many papers and watched so many videos, I now am at least aware of its existence and have started familiarizing myself with some of the jargon.\n\nIf possible, could you share a bit about what will happen after the competition ends? For example, what will the winning entries of this competition be used for exactly? The [WIFS website](https://www.wifs2020.nyu.edu/home) is pretty vague and essentially purports that the goal here is simply to introduce / involve the greater ML community to steganography and steganalysis, and to expand the boundaries of state-of-the-art.\n\nI stem from the applied DS side (I'm a contractor, not a researcher!) so whenever I see money on the table, there's usually some party pushing an agenda or trying to achieve some measurable objective. I don't know how ALASKA2 was funded, and it's not like I'm entitled to that information. But as I discover more about steganalysis specifically, I'm also curious how the results of this effort will be applied, who the groups organizing this effort are, and what they're all about. Since you are very deep in that world, any additional information you can share on that front would be greatly appreciated.\n\nThank you!", "914226": "This is a question for Remi but I will take a stab. My understanding is that the main goal of this competition is to push the state of the art in steganalysis in more real-life conditions, e.g., for current best stego schemes that are not necessarily available to the steganalyst, variable and unknown payload size, and when the embedding is spread among color channels. The detectors we build (will have built) are unlikely to be used in practice IMHO as only three embedding schemes are involved. There are dozens of steganographic schemes, many old ones, that are in fact much more likely to be used in practice by steganography users, such as F5, OutGuess, JP Hide&Seek, Steghide, Jsteg, etc, etc. Plus, this competition does not address the problem of how to steganalyze larger images. What if you have 20 megapixel image, for example? How would you make a decision about it using your detector? You should not resize as that would be highly suboptimal, so you may apply your detector to individual tiles of the image and \"fuse\" the decisions in some manner. This important aspect is completely missing here. Moreover, we only covered three quality factors. There are JPEGs with custom quantization tables, JPEGs with different chroma subsampling, and different encoding (RGB vs. YCrCb), etc. For a detector that could be used in real life, all this (immense) complexity would have to addressed.\n\nHaving said all this, this competition has already brought a lot of fresh wind to the field of steganography that will lead to many new publications when properly studied in a controlled environment. Remember, academicians want understanding, they are far less interested in building actual detectors. So, to everyone competing here : \"great job, guys, you are helping our field to flourish by trying ideas that stego experts may not venture into.\" Apparently, engineering detectors is a job best left to deep learning experts who know just a little about steganography rather than stego experts who know a little about deep learning. This, by itself, is useful to know. Technology keeps on evolving, we retool, relearn, and embrace new paradigms, to ask more questions as we strive to understand. The better detectors will reshape the state of the art in steganography as yesterday's kings will fall and new kings emerge. The question of the best steganographic scheme will be revisited with this new generation of detectors. So far, it looks like the 7-year old J-UNIWARD may just survive the onslaught the best while the new kid on the block, a better theoretically founded J-MiPOD, may need to be retuned to be more secure in this never-ending game of cat and mouse. \n\nOnce more, thanks everyone for participating. It is much appreciated whoever wins by a hair width in the end." }, "source": "meta" }