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.
A nice tutorial (and a lot of articles) can be found on the Dartmouth site of Hany Farid.
http://www.cs.dartmouth.edu/farid/downloads/tutorials/digita...
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.