Too much false fake face detection

[Vadim2S]

For example: attached image. My cat with melon. Original photo, slightly unfocused. Please, ignore all text .

[CauchyComplete]

Hi, Thanks for your false-positive example report. Future works may focus more on images containing a relatively bright object compared to other regions, as your example suggests. Image manipulation detection is an ongoing research field and it is a very challenging task. I recommend you to compare the outputs with other state-of-the-art forgery detectors, such as ManTra-Net, Noiseprint, EXIF-SC. Public forensic dataset results say CAT-Net overperforms current best approaches. Thank you.

[Vadim2S]

Those project are dead. Python 3.5 and Tensorflow 1.8 required too much work just for launch. Cat-Net do good result, except:

1. "Magic wand" like tool removal. I am attach Spain picture with building crane at end of channel. In air it removed with "magic wand" but you can see it mirroring in water.
2. Shape alternation. I am attach Proportion picture with instagramm body alternation. You can see difference. Both cases do not detected at all .

[CauchyComplete]

Thanks for your good analysis. However, please keep in mind that CAT-Net targeted splicing and copy-move forgeries. You know, CAT-Net mainly uses compression artifacts to detect splicings or copy-moves since pasted regions will likely be misaligned with the original image. Object removal like 'magic wand' and shape alternation forgeries are indeed very popular manipulation in real-world scenarios, but those types of forgeries were not used to train CAT-Net. This was mainly due to a lack of those types of datasets. Building a network that can detect those manipulations jointly with copy-pasting would be nice work though. ☺️

[dan326326]

Those project are dead. Python 3.5 and Tensorflow 1.8 required too much work just for launch. Cat-Net do good result, except:

1. "Magic wand" like tool removal. I am attach Spain picture with building crane at end of channel. In air it removed with "magic wand" but you can see it mirroring in water.
2. Shape alternation. I am attach Proportion picture with instagramm body alternation. You can see difference. Both cases do not detected at all .

Hi, what data set are you using

[Vadim2S]

Sorry, it is my own small dataset with 5-10 typical fake examples for each type alteration (insert, delete, alter) and each type instrument (Paint, Photoshop, Neural-Net). Nothing private but too unorganized for share.

[dan326326]

I see. Thank you

------------------ 原始邮件 ------------------ 发件人: "mjkwon2021/CAT-Net" @.>; 发送时间: 2021年11月10日(星期三) 晚上7:47 @.>; @.@.>; 主题: Re: [mjkwon2021/CAT-Net] Too much false fake face detection (#3)

Sorry, it is my own small dataset with 5-10 typical fake examples for each type alteration (insert, delete, alter) and each type instrument (Paint, Photoshop, Neural-Net). Nothing private but too unorganized for share.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.