Image Forgery Detection — Classical DIP on CASIA v2.0

Finding 1 · the dataset

The files give the answer away

0.990

balanced accuracy from file metadata alone on CASIA v2.0 as released (AUC 0.999). Most tampered images are TIFF files; authentic images are JPEGs, 89 % of them with standard IJG tables that no tampered JPEG uses.

Protocol	All shortcuts
A · files as released	0.989
C · JPEG files only	0.986
B · re-encoded as JPEG Q85	0.797
R · downscaled ×0.75, then re-encoded (main)	0.649

Balanced accuracy of classifiers that see only global cues (metadata, size, compression statistics). A forensic feature counts only if it beats these under the same protocol.

Bar chart: balanced accuracy of shortcut features under protocols A, C, B and R — outputs/10_leakage_shortcuts.png

The pipeline, in order

Five course topics, eight techniques

Every step on one validation copy-move image (the pasted person on the right is a mirrored copy of the one on the left), after protocol R. Each module compares a region with the rest of its own image.

0 · input

Protocol R and ground truth

The image is downscaled ×0.75 and re-saved as JPEG Q85 before any analysis, which erases file-format and compression-history shortcuts.

Ground-truth region — 01b ground truth (green)

1 · point processing

Error level analysis

|I − JPEG_q(I)| per pixel: a region with a different compression history re-compresses differently. Divided by block texture so busy areas are not mistaken for tampering.

Texture-normalised ELA — 02b texture-normalised

2 · histogram processing

Equalisation and χ² patch histograms

The low-contrast ELA residual is histogram-equalised; each patch's histogram is compared with the image's typical one.

Histograms before and after equalisation — 03b before / after

3 · spatial filtering

Local noise level

A high-pass kernel leaves only sensor noise; its robust σ is estimated per block over non-edge pixels. A region from another camera often carries different noise.

4 · frequency domain

DCT double quantisation and JPEG ghosts

JPEG quantises 8×8 DCT coefficients, so their histogram spikes at multiples of the step; double compression leaves a periodic pattern that pasted regions lack. JPEG ghosts re-save at many qualities and look for regions that disagree.

DCT coefficient histogram — 05a DCT histogram

5 · edge & corner detection

Canny sharpness and SIFT copy-move

Canny edges measure boundary sharpness. SIFT keypoints are matched against the image itself and its mirror, then verified with RANSAC — here it finds both the source and the flipped copy (14 inliers).

Copy-move keypoint evidence — 06c keypoint copy-move

6 · detection

Fusion, verdict and region

A calibrated random forest fuses 82 features into P(tampered) and abstains inside an uncertain band; gradient boosting fuses the eight maps into a mask. Verdict here: tampered (P = 0.99), likely copy-move.

Fused tampering probability — 07a fused probability

Detected region overlay — 07b suspected region

Every number, read from the run

Test set, evaluated once

1,892 held-out images, frozen by hash before any model was trained, evaluated once; every later use is logged. Protocol R unless stated.

0.805

test AUC, 95 % CI [0.779, 0.831], against 0.712 for the shortcuts

+0.103

added over the shortcut features, CI [0.076, 0.129]

86 %

accuracy on the 52 % of images the calibrated detector judges

0.972

AUC on MICC-F220, an unseen copy-move dataset (11 forged scenes)

At threshold 0.5	R	B
Accuracy	0.749	0.850
Precision	0.756	0.813
Recall	0.567	0.819
F1	0.648	0.816
Localisation pixel F1	0.198	0.308

Deployed detector (R)	authentic	tampered	uncertain
truly authentic (1,123)	619	13	491
truly tampered (769)	126	222	421

Robust to re-saving: AUC ≥ 0.78 down to JPEG quality 50; a social-media-style upload costs 0.04. Main weaknesses: same-history splices, low-detail images, self-similar scenes (all 13 false alarms had copy-move matches).

Test ROC curves under protocol R — outputs/15_test_roc.png

Finding 2 · deep learning

An ELA-CNN scores highest only on the leaking files

A pretrained ResNet-18 on ELA images, same splits and protocols, pre-registered and evaluated once.

Test AUC	ELA-CNN	Classical
A · files as released	0.994	0.981
B · re-encoded	0.811	0.920
R · strict	0.770	0.805
MICC-F220 (R)	0.529	0.972

On the released files the shortcut features alone reach 1.000. Under R the CNN trails by 0.036 [0.005, 0.067] and is not significantly better than a shortcut SVM; on MICC-F220 its AUC is consistent with chance.

ELA of an authentic JPEG and a tampered TIFF, as released and after protocol R — ELA as released: the authentic JPEG is almost black, the tampered TIFF lights up

Validation, including what went wrong

Mistakes caught before the test set was touched

DCT periodicity test: 16–48 % false positives on wide histograms → replaced by a per-histogram permutation null (0–3 %).
Untuned defaults: the detector entry point ignored the tuned parameters → it now refuses a model whose recorded parameters differ.
Copy-move: mirrored copies were missed and unoriented pairs split one copy into two → mirror matching and pair orientation.
Noise: flat patches became false low-noise outliers → excluded per pixel.
A pre-registered check that was invalid: the CNN size check broke ties by label order → withdrawn, replaced by a labelled post-hoc analysis.
Test-set hygiene: every use of the test split is logged with file fingerprints, including two early runs disclosed afterwards.

209 automated tests cover the modules on synthetic forgeries with known regions, the protocols, the split invariants, the statistics and the interface.

Run it

From a fresh clone

git clone https://github.com/Smit2102/image-forgery-detection.git
cd image-forgery-detection
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
# place CASIA v2.0 at CASIA2/, then run the pipeline scripts in order (see README)
.venv/bin/python detect.py photo.jpg          # one image -> verdict, mask, overlay, report.json
.venv/bin/streamlit run app/Home.py           # interactive tool at http://localhost:8510

The whole pipeline on one image — outputs/08_summary_grid.png — the whole pipeline on one image