Anyone tried this as a multiclass classification with 75000 sample for each class ? Did it do better than binary ?
", "messages": [ { "id": "835201", "postDate": "05/06/2020 05:40:35", "content": "Anyone tried this as a multiclass classification with 75000 sample for each class ? Did it do better than binary ?
", "rawMarkdown": "Anyone tried this as a multiclass classification with 75000 sample for each class ? Did it do better than binary ?", "votes": null }, { "id": "835400", "postDate": "05/06/2020 08:45:03", "content": "Good question. Indeed, if you look at the team that won the ALASKA#1 challenge they have not merged all images together (cover and stego).\nThey have used a two criteria splits:\n1. Images compressed with different quality factors are separated and used to train different CNN.\n2. Different classifiers were used for each embedding scheme (from binary to multi-class the gain is about 2% accuracy)
\n\nNow, using the same trick for ALASKA2 would mean training 9 different classifiers, for 3 different quality factors and 3 different embedding scheme (or using a single multiclass classifier)
\n\nSee here for the paper: http://ws.binghamton.edu/fridrich/Research/ALASKA-preprint1.pdf and especially beginning of section 2.3 and Table1 for your question.\nand there for the code: https://github.com/YassineYousfi/alaska
", "rawMarkdown": "Good question. Indeed, if you look at the team that won the ALASKA#1 challenge they have not merged all images together (cover and stego).\nThey have used a two criteria splits:\n1. Images compressed with different quality factors are separated and used to train different CNN.\n2. Different classifiers were used for each embedding scheme (from binary to multi-class the gain is about 2% accuracy)\n\nNow, using the same trick for ALASKA2 would mean training 9 different classifiers, for 3 different quality factors and 3 different embedding scheme (or using a single multiclass classifier)\n\nSee here for the paper: [http://ws.binghamton.edu/fridrich/Research/ALASKA-preprint1.pdf](http://ws.binghamton.edu/fridrich/Research/ALASKA-preprint1.pdf) and especially beginning of section 2.3 and Table1 for your question.\nand there for the code: [https://github.com/YassineYousfi/alaska](https://github.com/YassineYousfi/alaska)", "votes": null }, { "id": "835404", "postDate": "05/06/2020 08:51:23", "content": "Thanks a lot for the reference .
", "rawMarkdown": "Thanks a lot for the reference .", "votes": null }, { "id": "869196", "postDate": "05/31/2020 20:07:29", "content": "YES! I was trying to figure out WHYYYY a notebook was using 9 classes. I was so confused, where were they coming from? I used 4 (Cover + 3x Embedding Schemes), but now I finally understand the 9. Thank you so much for your post! 🙏
", "rawMarkdown": "YES! I was trying to figure out WHYYYY a notebook was using 9 classes. I was so confused, where were they coming from? I used 4 (Cover + 3x Embedding Schemes), but now I finally understand the 9. Thank you so much for your post! 🙏", "votes": null } ], "comments": [ { "id": 835400, "author_name": "remicogranne", "author_url": "", "post_date": "05/06/2020 08:45:03", "content": "Good question. Indeed, if you look at the team that won the ALASKA#1 challenge they have not merged all images together (cover and stego).\nThey have used a two criteria splits:\n1. Images compressed with different quality factors are separated and used to train different CNN.\n2. Different classifiers were used for each embedding scheme (from binary to multi-class the gain is about 2% accuracy)
\n\nNow, using the same trick for ALASKA2 would mean training 9 different classifiers, for 3 different quality factors and 3 different embedding scheme (or using a single multiclass classifier)
\n\nSee here for the paper: http://ws.binghamton.edu/fridrich/Research/ALASKA-preprint1.pdf and especially beginning of section 2.3 and Table1 for your question.\nand there for the code: https://github.com/YassineYousfi/alaska
", "votes": null, "replies": [ { "id": 835404, "author_name": "phoenix9032", "author_url": "", "post_date": "05/06/2020 08:51:23", "content": "Thanks a lot for the reference .
", "votes": null, "replies": [] } ] }, { "id": 869196, "author_name": "andradaolteanu", "author_url": "", "post_date": "05/31/2020 20:07:29", "content": "YES! I was trying to figure out WHYYYY a notebook was using 9 classes. I was so confused, where were they coming from? I used 4 (Cover + 3x Embedding Schemes), but now I finally understand the 9. Thank you so much for your post! 🙏
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "835201": "Anyone tried this as a multiclass classification with 75000 sample for each class ? Did it do better than binary ?", "835400": "Good question. Indeed, if you look at the team that won the ALASKA#1 challenge they have not merged all images together (cover and stego).\nThey have used a two criteria splits:\n1. Images compressed with different quality factors are separated and used to train different CNN.\n2. Different classifiers were used for each embedding scheme (from binary to multi-class the gain is about 2% accuracy)\n\nNow, using the same trick for ALASKA2 would mean training 9 different classifiers, for 3 different quality factors and 3 different embedding scheme (or using a single multiclass classifier)\n\nSee here for the paper: [http://ws.binghamton.edu/fridrich/Research/ALASKA-preprint1.pdf](http://ws.binghamton.edu/fridrich/Research/ALASKA-preprint1.pdf) and especially beginning of section 2.3 and Table1 for your question.\nand there for the code: [https://github.com/YassineYousfi/alaska](https://github.com/YassineYousfi/alaska)", "835404": "Thanks a lot for the reference .", "869196": "YES! I was trying to figure out WHYYYY a notebook was using 9 classes. I was so confused, where were they coming from? I used 4 (Cover + 3x Embedding Schemes), but now I finally understand the 9. Thank you so much for your post! 🙏" }, "source": "meta" }