![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F24375367%2F6549a833598da60a2324ea2e2789ce34%2F90.contour_result.png?generation=1761710014802615&alt=media\")
The forged directory contains 374 images that have no corresponding images in the authentic directory. I looked at the masks for these images, many of them do not seem correct. They don't overlap duplicated regions. Some of the examples are listed here. The ids are 90, 9822, 1070, 13734, 1661, from top to bottom.
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F24375367%2Fb372a48da8e631b9faf9fcc9225a09b1%2F9822.contour_result.png?generation=1761710060629232&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F24375367%2Fdf07cbd62dddba6bf73d1ebde03303d1%2F1070.contour_result.png?generation=1761710105693063&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F24375367%2Fbd1433e52f22d685ea4297f4ed4f9616%2F13734.contour_result.png?generation=1761710141022883&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F24375367%2Fb86abb0c11ebc9ac2374661a2c74a3b8%2F1661.contour_result.png?generation=1761710454418200&alt=media\")
The forged directory contains 374 images that have no corresponding images in the authentic directory. I looked at the masks for these images, many of them do not seem correct. They don't overlap duplicated regions. Some of the examples are listed here. The ids are 90, 9822, 1070, 13734, 1661, from top to bottom.
\n
nice catch !
", "rawMarkdown": "nice catch !", "votes": null }, { "id": "3308439", "postDate": "10/29/2025 12:33:36", "content": "Thanks for your comment!
\nThese masks are correct. Note that in scientific images, forgeries can occur to duplicate an object, or (as in these cases) to hide an object by pasting a piece of the background over it. That is what is occurring in the examples you presented.
\nCheck the difference between the authentic and forged images to understand what is happening.\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11373480%2F151df79a678026afe02902ece04bd009%2Fbackground-forgery.png?generation=1761741184566797&alt=media\")
Here, the part of the image that is copied seems to have a lot of different possible origins. Am I wrong?
\nhere's an example of what I mean
\nSince there can be many different possibilities, how are we (and the AI model we're building) supposed to know that the one you attached is the correct mask and what I have here is not correct?
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5956875%2F3206426043abdb5f663ad91ef250d941%2Finbox_11373480_151df79a678026afe02902ece04bd009_background-forgery.png?generation=1762630371314141&alt=media\")
Thanks for the question.
\nIn forensics, we use the Locard's exchange principle, which states: \"Every contact leaves a trace.\" So, there is almost always a way to catch the forgery.
\nIn this type of forgery, you should also consider the image noise surrounding the forged regions, which might not be easily noticeable without processing the image.
\nFor example, if you modify the color curves, you will often see that the same noise patterns occur in the source region and the copied regions. This allows us to verify that a region has been copied from one location to another.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F11373480%2F34c9647262711972cd65240743dcdcd8%2Finbox_11373480_151df79a678026afe02902ece04bd009_background-forgery.png?generation=1762781159315144&alt=media\")
This is just one example of how to detect this, and there are surely many other methods that are more robust to the problem :)
", "rawMarkdown": "Thanks for the question.\n\nIn forensics, we use the Locard's exchange principle, which states: \"Every contact leaves a trace.\" So, there is almost always a way to catch the forgery.\n\nIn this type of forgery, you should also consider the image noise surrounding the forged regions, which might not be easily noticeable without processing the image.\n\nFor example, if you modify the color curves, you will often see that the same noise patterns occur in the source region and the copied regions. This allows us to verify that a region has been copied from one location to another.\n\n\n\nThis is just one example of how to detect this, and there are surely many other methods that are more robust to the problem :)", "votes": null } ], "comments": [ { "id": 3308347, "author_name": "steubk", "author_url": "", "post_date": "10/29/2025 07:07:38", "content": "nice catch !
", "votes": null, "replies": [] }, { "id": 3308439, "author_name": "joophillipecardenuto", "author_url": "", "post_date": "10/29/2025 12:33:36", "content": "Thanks for your comment!
\nThese masks are correct. Note that in scientific images, forgeries can occur to duplicate an object, or (as in these cases) to hide an object by pasting a piece of the background over it. That is what is occurring in the examples you presented.
\nCheck the difference between the authentic and forged images to understand what is happening.\n
Here, the part of the image that is copied seems to have a lot of different possible origins. Am I wrong?
\nhere's an example of what I mean
\nSince there can be many different possibilities, how are we (and the AI model we're building) supposed to know that the one you attached is the correct mask and what I have here is not correct?
\n
Thanks for the question.
\nIn forensics, we use the Locard's exchange principle, which states: \"Every contact leaves a trace.\" So, there is almost always a way to catch the forgery.
\nIn this type of forgery, you should also consider the image noise surrounding the forged regions, which might not be easily noticeable without processing the image.
\nFor example, if you modify the color curves, you will often see that the same noise patterns occur in the source region and the copied regions. This allows us to verify that a region has been copied from one location to another.
\n
This is just one example of how to detect this, and there are surely many other methods that are more robust to the problem :)
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3308306": "The forged directory contains 374 images that have no corresponding images in the authentic directory. I looked at the masks for these images, many of them do not seem correct. They don't overlap duplicated regions. Some of the examples are listed here. The ids are 90, 9822, 1070, 13734, 1661, from top to bottom.\n\n\n\n\n\n\n\n\n\n", "3308347": "nice catch !", "3308439": "Thanks for your comment!\n\nThese masks are correct. Note that in scientific images, forgeries can occur to duplicate an object, or (as in these cases) **to hide an object by pasting a piece of the background over it**. That is what is occurring in the examples you presented.\n\nCheck the difference between the authentic and forged images to understand what is happening.\n", "3313160": "Here, the part of the image that is copied seems to have a lot of different possible origins. Am I wrong?\n\nhere's an example of what I mean\n\nSince there can be many different possibilities, how are we (and the AI model we're building) supposed to know that the one you attached is the correct mask and what I have here is not correct?\n\n", "3315075": "Thanks for the question.\n\nIn forensics, we use the Locard's exchange principle, which states: \"Every contact leaves a trace.\" So, there is almost always a way to catch the forgery.\n\nIn this type of forgery, you should also consider the image noise surrounding the forged regions, which might not be easily noticeable without processing the image.\n\nFor example, if you modify the color curves, you will often see that the same noise patterns occur in the source region and the copied regions. This allows us to verify that a region has been copied from one location to another.\n\n\n\nThis is just one example of how to detect this, and there are surely many other methods that are more robust to the problem :)" }, "source": "meta" }