Hacker News Reader

HNHacker News

Top New Best Ask Show Jobs

Comment by devoply | Hacker News Reader

Parent Full thread

devoply·

https://fotoforensics.com/analysis.php?id=8497eb5680338dbc87...

https://fotoforensics.com/analysis.php?id=675fe6f23e33beb5fd...

https://fotoforensics.com/analysis.php?id=cde26cacf7826ac00e...

View on HN

Except, if you've looked at enough of these, you know that they are often very misleading, especially to the untrained eye. It's a fantastic tool for people to pretend that they have "proof", but all it takes is showing the ELA of a photo you know to be real to realize that the tool is more complex that it appears at first.

Here's two taken from NatGeo's best photo's of 2018 [0]:

https://fotoforensics.com/analysis.php?id=a29db81cd8bfc2e23a...

https://fotoforensics.com/analysis.php?id=7a9b396c6b9ee467d3...

To make things worse, you used the low-res/low-jpeg-quality version of the photo from the blog, which inherently have more error. If you use the original 2100x3100 photo [1], you get a much different result:

https://fotoforensics.com/analysis.php?id=c6c3a0784144868a56...

[0]: https://www.nationalgeographic.com/photography/best-of-2018/...

[1] http://time.com/3992951/iss-photobomb/

Could you elaborate on meaning of these?

They probably show a JPEG diff: The quantization errors introduced by JPEG-compression. This is relevant to the "fake" discussion because often the source material of shopped images is JPEG compressed and has JPEG-artifacts. If one combines source material of different shots their JPEG artifacts will likely not align because of the different resolution, offset, and compression levels of the parts. This means that the inserted parts of an image will have a different noise-profile which can often be spotted just by looking at the diff.

I'm not trained to read these diffs and somebody who's cheating professionally would know to cover their tracks. So while they can be used to detect forgery they cannot disprove it.

Looks like it's showing this: https://en.wikipedia.org/wiki/Error_level_analysis

Reply on news.ycombinator.com