This competition is quite different from others in the sense that it is not only based on classical deep learning methods but also on steganography analysis which opens a few more doors in the approaches taken in cracking the problem.
\n\nSome of them include:
\n\n1) Classic deep learning models for predictions proposed by\n @xhlulu Notebook (ENet Model)\n@tanulsingh077 Notebook (Conv Model)
\n\nMy inferences:\n- Performs perfectly with accuracies as high as 99% on public data.\n- Possibility of failing severely if the private data is from an entirely different distribution (Considering the possibility that the hidden message length can be altered in the images).
\n\n2) JPEG Compression analysis proposed by \n@miklgr500 Notebook
\n\nMy inferences:\n- Model performs beyond expectations considering all we do is obtain results from the compression rate (86% on public data).\n- Considerable room for improvement by arriving upon that perfect function based on the compression rate. My forked notebook here contains some of the functions that I experimented upon.\n- Almost certain to fail on real world data where the compression rates can be widely varying and often misleading.
\n\nMore methods will be appended to this list along with suitable inferences when I come across them.
\n\nFeel free to comment below with new approaches or inferences!
", "messages": [ { "id": "825926", "postDate": "04/29/2020 10:33:40", "content": "This competition is quite different from others in the sense that it is not only based on classical deep learning methods but also on steganography analysis which opens a few more doors in the approaches taken in cracking the problem.
\n\nSome of them include:
\n\n1) Classic deep learning models for predictions proposed by\n @xhlulu Notebook (ENet Model)\n@tanulsingh077 Notebook (Conv Model)
\n\nMy inferences:\n- Performs perfectly with accuracies as high as 99% on public data.\n- Possibility of failing severely if the private data is from an entirely different distribution (Considering the possibility that the hidden message length can be altered in the images).
\n\n2) JPEG Compression analysis proposed by \n@miklgr500 Notebook
\n\nMy inferences:\n- Model performs beyond expectations considering all we do is obtain results from the compression rate (86% on public data).\n- Considerable room for improvement by arriving upon that perfect function based on the compression rate. My forked notebook here contains some of the functions that I experimented upon.\n- Almost certain to fail on real world data where the compression rates can be widely varying and often misleading.
\n\nMore methods will be appended to this list along with suitable inferences when I come across them.
\n\nFeel free to comment below with new approaches or inferences!
", "rawMarkdown": "This competition is quite different from others in the sense that it is not only based on classical deep learning methods but also on steganography analysis which opens a few more doors in the approaches taken in cracking the problem.\n\nSome of them include:\n\n1) Classic deep learning models for predictions proposed by\n @xhlulu [Notebook](https://www.kaggle.com/xhlulu/alaska2-efficientnet-on-tpus) (ENet Model)\n@tanulsingh077 [Notebook](https://www.kaggle.com/tanulsingh077/steganalysis-approaching-as-regression-problem) (Conv Model)\n\nMy inferences:\n- Performs perfectly with accuracies as high as 99% on public data.\n- Possibility of failing severely if the private data is from an entirely different distribution (Considering the possibility that the hidden message length can be altered in the images).\n\n\n2) JPEG Compression analysis proposed by \n@miklgr500 [Notebook](https://www.kaggle.com/miklgr500/jpeg-compression-rate)\n\nMy inferences:\n- Model performs beyond expectations considering all we do is obtain results from the compression rate (86% on public data).\n- Considerable room for improvement by arriving upon that perfect function based on the compression rate. My forked notebook [here](https://www.kaggle.com/akashsuper2000/jpeg-compression-rate?scriptVersionId=32883953) contains some of the functions that I experimented upon.\n- Almost certain to fail on real world data where the compression rates can be widely varying and often misleading.\n\nMore methods will be appended to this list along with suitable inferences when I come across them.\n\nFeel free to comment below with new approaches or inferences!", "votes": null }, { "id": "846734", "postDate": "05/14/2020 01:05:26", "content": "Thanks for this thread. I will be following it intently for future updates. It is a good jumping off point for me.
\n\nCheers
", "rawMarkdown": "Thanks for this thread. I will be following it intently for future updates. It is a good jumping off point for me.\n\nCheers", "votes": null } ], "comments": [ { "id": 846734, "author_name": "stevenknight", "author_url": "", "post_date": "05/14/2020 01:05:26", "content": "Thanks for this thread. I will be following it intently for future updates. It is a good jumping off point for me.
\n\nCheers
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "825926": "This competition is quite different from others in the sense that it is not only based on classical deep learning methods but also on steganography analysis which opens a few more doors in the approaches taken in cracking the problem.\n\nSome of them include:\n\n1) Classic deep learning models for predictions proposed by\n @xhlulu [Notebook](https://www.kaggle.com/xhlulu/alaska2-efficientnet-on-tpus) (ENet Model)\n@tanulsingh077 [Notebook](https://www.kaggle.com/tanulsingh077/steganalysis-approaching-as-regression-problem) (Conv Model)\n\nMy inferences:\n- Performs perfectly with accuracies as high as 99% on public data.\n- Possibility of failing severely if the private data is from an entirely different distribution (Considering the possibility that the hidden message length can be altered in the images).\n\n\n2) JPEG Compression analysis proposed by \n@miklgr500 [Notebook](https://www.kaggle.com/miklgr500/jpeg-compression-rate)\n\nMy inferences:\n- Model performs beyond expectations considering all we do is obtain results from the compression rate (86% on public data).\n- Considerable room for improvement by arriving upon that perfect function based on the compression rate. My forked notebook [here](https://www.kaggle.com/akashsuper2000/jpeg-compression-rate?scriptVersionId=32883953) contains some of the functions that I experimented upon.\n- Almost certain to fail on real world data where the compression rates can be widely varying and often misleading.\n\nMore methods will be appended to this list along with suitable inferences when I come across them.\n\nFeel free to comment below with new approaches or inferences!", "846734": "Thanks for this thread. I will be following it intently for future updates. It is a good jumping off point for me.\n\nCheers" }, "source": "meta" }