All cameras introduce varying degrees of imperfections. While these imperfections are undesirable for the consumer, they can be highly informative in an image forensic setting. Sensor imperfections, thermal effects, compression, etc. each introduce noise in the image. This noise manifests itself as slight variations in the recorded intensities of light. If a portion of an image is altered then it is likely that the underlying noise pattern will be disrupted. Differences in the noise pattern can therefore be used to detect and localize image tampering.
One of my former graduate students, Professor Siwei Lyu, recently developed a very nice forensic technique for measuring and detecting inconsistencies in image noise. This is a particularly challenging problem because estimating noise is a highly unconstrained problem. By exploiting statistical regularities in natural images, the problem of noise estimation can be more constrained. And most impressively, this estimate can be performed on relatively small image blocks.
Shown below, for example, is a (obviously) fake image and the results of forensically analyzing the image noise. The white pixels on the right correspond to regions with different amounts of noise relative to the rest of the image.
Although the technical details are a bit involved, the actual estimate is fairly straight forward and computationally efficient (see “Exposing Image Splicing with Inconsistent Local Noise Variances” for more details). As with all forensic techniques, there are limitations to this analysis. The actual image content can impact the estimate of noise and lead to false positives. And, a sufficiently informed forger can remove the existing noise and re-introduce a consistent noise pattern. Nevertheless, this technique, when coupled with other analyses should prove to be a useful tool in your forensic toolbox.