BitMix: Data Augmentation for Image Steganalysis\n\nConvolutional neural networks (CNN) for image steganalysis demonstrate better performances with employing concepts from highlevel vision tasks. The major employed concept is to use data\naugmentation to avoid overfitting due to limited data. To augment data without damaging the message embedding, only rotating multiples of 90◦ or horizontally flipping are used in steganalysis, which generates eight fixed results from one sample. To overcome this limitation,\nwe propose BitMix, a data augmentation method for spatial image steganalysis. BitMix mixes a cover and stego image pair by swapping the random patch and generates an embedding adaptive label with the ratio of the number of pixels modified in the swapped patch to those\nin the cover-stego pair. We explore optimal hyperparameters, the ratio of applying BitMix in the mini-batch, and the size of the bounding box for swapping patch. The results reveal that using BitMix improves the performance of spatial image steganalysis and better than other data\naugmentation methods.\n
BitMix: Data Augmentation for Image Steganalysis\n\nConvolutional neural networks (CNN) for image steganalysis demonstrate better performances with employing concepts from highlevel vision tasks. The major employed concept is to use data\naugmentation to avoid overfitting due to limited data. To augment data without damaging the message embedding, only rotating multiples of 90◦ or horizontally flipping are used in steganalysis, which generates eight fixed results from one sample. To overcome this limitation,\nwe propose BitMix, a data augmentation method for spatial image steganalysis. BitMix mixes a cover and stego image pair by swapping the random patch and generates an embedding adaptive label with the ratio of the number of pixels modified in the swapped patch to those\nin the cover-stego pair. We explore optimal hyperparameters, the ratio of applying BitMix in the mini-batch, and the size of the bounding box for swapping patch. The results reveal that using BitMix improves the performance of spatial image steganalysis and better than other data\naugmentation methods.\n
thanks @bibek777 any link for implementation thanks
", "rawMarkdown": "thanks @bibek777 any link for implementation thanks", "votes": null }, { "id": "910683", "postDate": "07/01/2020 09:57:40", "content": "This is similar to cutmix augmentation I guess ? I've just heard about it during Chris Deotte presentation last week\nhttps://www.kaggle.com/cdeotte/how-to-compete-with-gpus-workshop#STEP-2:-Data-Augmentation
", "rawMarkdown": "This is similar to cutmix augmentation I guess ? I've just heard about it during Chris Deotte presentation last week\nhttps://www.kaggle.com/cdeotte/how-to-compete-with-gpus-workshop#STEP-2:-Data-Augmentation", "votes": null }, { "id": "910730", "postDate": "07/01/2020 10:25:27", "content": "great @nyleve just a quick one are you training on kaggle gpu or tpu
", "rawMarkdown": "great @nyleve just a quick one are you training on kaggle gpu or tpu", "votes": null }, { "id": "911223", "postDate": "07/01/2020 16:04:01", "content": "GPU, but working on TPU at the moment since I'm out of quota
", "rawMarkdown": "GPU, but working on TPU at the moment since I'm out of quota", "votes": null }, { "id": "911805", "postDate": "07/02/2020 03:12:44", "content": "Hey, any reference links would help us all a lot.
", "rawMarkdown": "Hey, any reference links would help us all a lot.", "votes": null }, { "id": "912440", "postDate": "07/02/2020 13:26:47", "content": "\n\n\nTo augment data without damaging the message embedding, only rotating multiples of 90◦ or horizontally flipping are used in steganalysis, which generates eight fixed results from one sample
\n
It's just declaring.\nDCT coefficients is symmetric or assymetric, so if you make Horizontal or Vertical Flip then you just changing the sign for assymetric coeffs. \nIf you rotate an image then you switch rows and columns, but JPEG's quantization matrix is not rotation-invariant.\nTest 3021:\nDefine Quantization Table 0 precision 0\n 3 2 2 3 5 8 10 12\n 2 2 3 4 5 12 12 11\n 3 3 3 5 8 11 14 11\n 3 3 4 6 10 17 16 12\n 4 4 7 11 14 22 21 15\n 5 7 11 13 16 21 23 18\n 10 13 16 17 21 24 24 20\n 14 18 19 20 22 20 21 20\nSo non-zero coeffs can become zeros.
", "rawMarkdown": "> To augment data without damaging the message embedding, only rotating multiples of 90◦ or horizontally flipping are used in steganalysis, which generates eight fixed results from one sample\n> \n\n\nIt's just declaring.\nDCT coefficients is symmetric or assymetric, so if you make Horizontal or Vertical Flip then you just changing the sign for assymetric coeffs. \nIf you rotate an image then you switch rows and columns, but JPEG's quantization matrix is not rotation-invariant.\nTest 3021:\nDefine Quantization Table 0 precision 0\n 3 2 2 3 5 8 10 12\n 2 2 3 4 5 12 12 11\n 3 3 3 5 8 11 14 11\n 3 3 4 6 10 17 16 12\n 4 4 7 11 14 22 21 15\n 5 7 11 13 16 21 23 18\n 10 13 16 17 21 24 24 20\n 14 18 19 20 22 20 21 20\nSo non-zero coeffs can become zeros.", "votes": null } ], "comments": [ { "id": 910458, "author_name": "pranshu29", "author_url": "", "post_date": "07/01/2020 06:50:11", "content": "thanks @bibek777 any link for implementation thanks
", "votes": null, "replies": [] }, { "id": 910683, "author_name": "nyleve", "author_url": "", "post_date": "07/01/2020 09:57:40", "content": "This is similar to cutmix augmentation I guess ? I've just heard about it during Chris Deotte presentation last week\nhttps://www.kaggle.com/cdeotte/how-to-compete-with-gpus-workshop#STEP-2:-Data-Augmentation
", "votes": null, "replies": [ { "id": 910730, "author_name": "pranshu29", "author_url": "", "post_date": "07/01/2020 10:25:27", "content": "great @nyleve just a quick one are you training on kaggle gpu or tpu
", "votes": null, "replies": [] }, { "id": 911223, "author_name": "nyleve", "author_url": "", "post_date": "07/01/2020 16:04:01", "content": "GPU, but working on TPU at the moment since I'm out of quota
", "votes": null, "replies": [] } ] }, { "id": 911805, "author_name": "akshat0007", "author_url": "", "post_date": "07/02/2020 03:12:44", "content": "Hey, any reference links would help us all a lot.
", "votes": null, "replies": [] }, { "id": 912440, "author_name": "demesgal", "author_url": "", "post_date": "07/02/2020 13:26:47", "content": "\n\n\nTo augment data without damaging the message embedding, only rotating multiples of 90◦ or horizontally flipping are used in steganalysis, which generates eight fixed results from one sample
\n
It's just declaring.\nDCT coefficients is symmetric or assymetric, so if you make Horizontal or Vertical Flip then you just changing the sign for assymetric coeffs. \nIf you rotate an image then you switch rows and columns, but JPEG's quantization matrix is not rotation-invariant.\nTest 3021:\nDefine Quantization Table 0 precision 0\n 3 2 2 3 5 8 10 12\n 2 2 3 4 5 12 12 11\n 3 3 3 5 8 11 14 11\n 3 3 4 6 10 17 16 12\n 4 4 7 11 14 22 21 15\n 5 7 11 13 16 21 23 18\n 10 13 16 17 21 24 24 20\n 14 18 19 20 22 20 21 20\nSo non-zero coeffs can become zeros.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "910045": "[BitMix: Data Augmentation for Image Steganalysis](https://arxiv.org/pdf/2006.16625.pdf)\n```\nConvolutional neural networks (CNN) for image steganalysis demonstrate better performances with employing concepts from highlevel vision tasks. The major employed concept is to use data\naugmentation to avoid overfitting due to limited data. To augment data without damaging the message embedding, only rotating multiples of 90◦ or horizontally flipping are used in steganalysis, which generates eight fixed results from one sample. To overcome this limitation,\nwe propose BitMix, a data augmentation method for spatial image steganalysis. BitMix mixes a cover and stego image pair by swapping the random patch and generates an embedding adaptive label with the ratio of the number of pixels modified in the swapped patch to those\nin the cover-stego pair. We explore optimal hyperparameters, the ratio of applying BitMix in the mini-batch, and the size of the bounding box for swapping patch. The results reveal that using BitMix improves the performance of spatial image steganalysis and better than other data\naugmentation methods.\n```", "910458": "thanks @bibek777 any link for implementation thanks", "910683": "This is similar to cutmix augmentation I guess ? I've just heard about it during Chris Deotte presentation last week\nhttps://www.kaggle.com/cdeotte/how-to-compete-with-gpus-workshop#STEP-2:-Data-Augmentation", "910730": "great @nyleve just a quick one are you training on kaggle gpu or tpu", "911223": "GPU, but working on TPU at the moment since I'm out of quota", "911805": "Hey, any reference links would help us all a lot.", "912440": "> To augment data without damaging the message embedding, only rotating multiples of 90◦ or horizontally flipping are used in steganalysis, which generates eight fixed results from one sample\n> \n\n\nIt's just declaring.\nDCT coefficients is symmetric or assymetric, so if you make Horizontal or Vertical Flip then you just changing the sign for assymetric coeffs. \nIf you rotate an image then you switch rows and columns, but JPEG's quantization matrix is not rotation-invariant.\nTest 3021:\nDefine Quantization Table 0 precision 0\n 3 2 2 3 5 8 10 12\n 2 2 3 4 5 12 12 11\n 3 3 3 5 8 11 14 11\n 3 3 4 6 10 17 16 12\n 4 4 7 11 14 22 21 15\n 5 7 11 13 16 21 23 18\n 10 13 16 17 21 24 24 20\n 14 18 19 20 22 20 21 20\nSo non-zero coeffs can become zeros." }, "source": "meta" }