2 H. Mareen et al.
1 Introduction
Manipulating digital images is both becoming easier and more realistic using
image editing tools and AI-based software. For example, Fig. 1a shows a manip-
ulated image in which the face of Captain Jack Sparrow is replaced with that of
another person (original from Pirates of the Caribbean). This forgery was gen-
erated in a few seconds using FaceHub.live, a free AI-based tool. Although there
are many harmless applications of these powerful tools, manipulated images con-
tribute to problems such as fake news, fake evidence, and fraud. Therefore, it
is crucial to investigate forensic methods that can fact check and verify images
found on the internet, i.e., in the wild.
Forensics forgery detection and localization methods typically look at im-
perceptible traces of an image’s digital history [23]. That is because every step
in the camera acquisition and digital editing process leaves a trail of clues. For
example, the sensor of the camera introduces a unique noise pattern [16], and
compression introduces blocking artifacts [3, 4, 12, 15, 27]. Although these traces
may be invisible to the human eye, a computer can exploit them. Unfortunately,
images found in the wild are often recompressed to a lower quality, which hides
these imperceptible traces. As such, successful forgery detection is challenging.
Recent methods that rely on deep Convolutional Neural Networks (CNNs)
have shown a higher level of robustness against recompression [22,23]. However,
most of them remain intrinsically weak because they can only detect forgeries
that were seen during training. Therefore, a promising strategy for forgery detec-
tion is using one-class deep-learning methods that are only trained using pristine
data. As such, they are not limited to specific types of forgery. For example, the
Noiseprint algorithm learns to extract a fingerprint from the camera model using
deep learning [7]. Then, inconsistencies in this fingerprint reveal tampering. This
is similar to traditional methods using PRNU fingerprints, but does not require
images from the camera of the image under investigation. It is trained on im-
ages coming from many different cameras, hence being able to detect different
sources, but no specific design was made during training to take into account
the different JPEG history of such patches.
In this paper, we exploit compression artifacts and extract a different finger-
print, called Comprint. It utilizes compression artifacts, but unlike most of the
state-of-the-art works, it does not assume that real regions underwent double
compression in contrast to fake regions that underwent single compression. In-
stead, it simply assumes that the pristine regions have a different compression
history than the tampered regions. Comprint is based on one-class training with
only pristine data: images compressed using a different Quality Factor (QF) or
quantization table in the JPEG coding standard. The architecture is based on a
Siamese network that is trained to distinguish regions that were compressed dif-
ferently. As such, the deep-learning method can extract a compression fingerprint
or comprint from an image under investigation. Then, a localization algorithm
segments the distinguished regions into a heatmap. For example, Fig. 1b and
Fig. 1c show the extracted comprint and heatmap, respectively, corresponding
to the manipulated image of Fig. 1a. We demonstrate that Comprint has a high