In more detail
Areas that were saved separately, or at a different quality, often show a different error level from the rest of the picture. ELA is only a hint: sharp edges, text and heavy saving can look suspicious in a genuine image, so it should be combined with other evidence. The method resaves the image at a fixed JPEG quality (often 90 or 95), subtracts it from the original pixel by pixel and amplifies the difference so it becomes visible. Areas that have already been through the same compression many times change very little; freshly inserted content changes more.
Why it matters
ELA is quick and needs nothing but the file, which makes it a popular first look in fact-checking. It is also easy to over-read: a bright patch can simply be high-contrast detail, and an image resaved several times or resized by a social network flattens all differences. Read it next to the EXIF data, the quantization tables and other tests such as PRNU.
Example
In a test made for this page, a 320×240 picture was saved twice at JPEG quality 75, then a 100×80 patch from the uncompressed source was pasted in and the result saved once more at 75. Resaving at 75 and measuring the average per-pixel difference (on a 0-255 scale):
| Area | Mean error |
|---|---|
| Pasted patch | 2.30 |
| Same area in the unedited image | 1.54 |
| Untouched area elsewhere | 1.34 |
The patch stands out, but only by about one level, which is why ELA output is shown amplified. Run it with the Image Forensics.