There is nothing scientific about any of these tests. What's a 'high' ELA or 'low'? The thing that triggers a 'positive' seems to be only the authors intuition. While some of the tests might show something, I'm pretty sure ELA is absolute garbage - you can't separate the number of needed resaves because of the content frequencies from number of times it has already been resaved.
One basically looks for roundings caused by quantization and places in the image where the compression (again quantization) is either inconsistent or non-optimal.
A nice tutorial (and a lot of articles) can be found on the Dartmouth site of Hany Farid.
Personally, I have had very very mixed results with this method, and never managed to model it correctly. Interpreting results always was a very human job.
I understand the theory and I certainly agree resaving jpegs will reduce the error level. If you have an undoctored image to compare it to, you could probably use that to determine which parts have been changed. But given only an image that may or may not have been doctored the error level will vary so much with the image content that it won't be meaningful.
For one, there should be a control. In this case, analyzing another ticket that has not been faked, but preferable he would not know if the control was real or not.
I think the author was clearly free from bias, and his results are easily reproduced by applying the same algorithms.
The subjectivity which you complain about may be the conclusions drawn from the results of the tests, which I think are distinct. Specifically, you cite his "intuition" as the origin of the conclusion.
I think the author's intuition is reliable because, like "real" scientists, he's an expert and speaks publicly about his work.[1] Or at least he appears to be. Are you prepared to challenge him as an expert?
I could, but shouldn't we demand he show some proof for his claims?
It seems to me that if you pull a bit of maths and technical magic out the normal skepticism the tech community melts away into a compliant bundle of gullibility.
While the result of the PCA is quite well defined, there is nothing well defined about the inferences he draws from that. In the end it shows a marked difference in the suspicious area, but there is no proof that the technique will only show those results on tampered areas.
If you can simply run filters until the suspicious area 'looks a bit different' and thats your success criteria, you haven't proven anything.
Neal specializes in this sort of thing. This blog post wasn't written to convince you that PCA (or ELA) work. If you've read his other posts on forensics, you'd see that he always starts with the common tests, then works through the most likely explanations, trying to see what you can rule out and where you should focus. A lot of the time, you can rule things out with common sense (the out-of-order numbers jumped out at me), though he always goes on to ferret out what manipulation was performed, because that's what the whole blog is about.
Anyhow, people have done lots of tests with lots of different tools to see what they report in different circumstances. So this isn't the first time that someone has used ELA or something. You will notice if you read past explanations that he always tries to find the exact manipulations done, rather than just running a tool and declaring something to have been Photoshopped.
I'm no expert, but I've read his blog for long enough to know that he knows an awful lot about the various quirks of many different image editing programs (Photoshop being by far the quirkiest). But if you still have questions, you can always email him. He was kind enough to reply with a lot of useful information when I asked him something a long time ago. Actually, I think he even blogged about it.
That's true, but trivially so. Things that aren't infallible can still be useful, but understanding their limitations is important.
If you read that blog, you can find plenty of discussion of those limitations, for example, how the absence of markers of digital manipulation does not prove that an image is genuine. After all, there are plenty of staged photos out there, images that were framed in a misleading way, etc.
I'm not making a trivial point though. I believe the tools, techniques and expertise this guy professes in image analysis are modern day snake oil. I just think that he has deluded himself instead of being deliberately dishonest.
How then do you explain the fact that his methods have worked? He has outed doctored photos that appeared in news stories and they have been retracted after investigation.
That's not quite what I'm saying, I don't think his tests are scientific not that they show absolutely nothing (although probably fairly close to nothing with ELA). I also haven't doubted his ability to detect fakes but I do doubt the particular efficacy and explanation of his techniques. There is no doubt using PCA can highlight unusual changes in an image - but there are also other explanations. His famous suggestion that terrorist videos had books inserted into them could simply have been a slightly different coloured spot light.
If his techniques are unreliable, he should have some notable failures by now after having analyzed so many images publicly. Where can these failures be found?
No not really, he techniques may be unhelpful and he still might be good at picking fakes. But analyzed so many images? There is a handful on the site and the outcome is almost never in doubt. You could test his ability to detect fakes but even if it were supreme you really couldn't separate that into the portion provided by his skills and those of his techniques.
Much more interesting would be to find out what the techniques really show, which faking techniques show which signatures and what other natural occurrences mimic that.
What about the times he found cheating by the winners of photographic contests? Those were hardly obvious choices. And falsely accusing someone would have really hurt his reputation. Of course, investigation proved that he was right. He analyzes something every few weeks it seems like (I've read the blog for years now), so yeah, there's a long history for you to look at. The outcome is only "never in doubt" if you're using hindsight bias. Besides, even when he already knows an image is fake, he figures out what is fake and how it was faked. One example would be showing which of those lottery numbers was real: the whole row was fake... except for the last number. Something which helped explain why the 2nd and 3rd photograph were the way that they were.
He's quite up front about the fact that some tests give inconclusive results. I believe that he has discussed the limitations of the tests. But even a test that gives an inconclusive result half the time can be useful if it shows you which areas of the image need attention the rest of the time. That data is indeed useful. He has talked about it. And you could just politely email him if you wanted to know more.
And yes, it probably is hard to separate the success due to technique and skills. But it's hard to believe that he would be good at picking out fake images without understanding why he was good at picking them out.
Well I enjoyed the discussion but I don't think I'm making the point especially well anymore, its not about picking fakes, its about reliably saying something about the numbers return by a mathematical function on an image. It needs more study and less experts making consulting money from their special magic.
