Image Forgery Detection using Error Level Analysis

Doctored photos travel faster than the truth: a spliced image can mislead thousands before anyone checks whether it is genuine. Journalists, fact-checkers and curious users need a practical way to inspect a suspicious JPEG — more rigorous than gut feeling but simpler than professional forensic software. Error level analysis fills that gap: each time a JPEG is saved, the whole image shares one compression history, but a region pasted from another photo usually carries a different one. Recompress at a known quality, measure the pixel-wise difference, and the pasted region lights up against the background. This project builds an ELA-based forgery analyzer on that principle — adjustable recompression quality, colorized heatmaps, connected-region detection and a confidence panel — evaluated honestly on the public CASIA v2 dataset. Its limits are part of the story: the project teaches exactly when ELA works, when it fails, and why it can never be called proof.

1. The uploaded JPEG is recompressed at a controlled quality setting and the pixel-wise difference (the error level) is computed.
2. The error map is amplified and colorized into an ELA heatmap, making recompression inconsistencies visible to the eye.
3. Connected-component analysis isolates regions whose error level deviates sharply from the image background.
4. Per-region statistics — mean error, size, deviation — feed a scoring rule that ranks regions by suspicion.
5. The demo app presents the original photo, the ELA heatmap and the flagged regions with their confidence scores.
6. The evaluation script runs the pipeline over CASIA v2 manipulated images and computes detection accuracy and false-positive rate during your build.

Tech stack:

• Python 3.10 (core ELA pipeline and analysis scripts)
• Pillow (JPEG recompression and error computation)
• OpenCV (heatmap rendering, connected components, overlays)
• NumPy (error statistics and region scoring)
• CASIA v2 dataset (public manipulated-image evaluation set)
• pandas (batch summary reports)
• Matplotlib (error distributions and evaluation plots)
• Streamlit demo app (photo upload, ELA heatmap, confidence panel)

• Genuine ELA heatmap generator: re-saves any JPEG at a configurable quality and renders the amplified pixel-wise difference as a forensic heatmap
• Adjustable analysis controls — re-save quality (70–98) and amplification factor — so the report can demonstrate how parameter choice affects visibility
• Block-level statistical feature extraction: error histograms, block variance and edge-density features computed over the ELA map
• XGBoost classifier trained on ELA feature vectors from the CASIA v2.0 benchmark, separating authentic from tampered images
• Batch folder scanning that processes whole image sets and writes a per-image tamper-score CSV report
• Flask demo UI with drag-and-drop upload, side-by-side original/ELA comparison and an adjustable tamper-score meter
• Evaluation notebook computing precision, recall, F1-score and ROC-AUC on your CASIA v2.0 split, with heatmap galleries of correct and missed cases

• Complete source code (ELA generator, feature extraction, classifier training, batch scanner, demo app)
• Jupyter training and evaluation notebook (buyer-run procedure: prepare CASIA v2.0 split, train, evaluate)
• Project report PDF (forensics background, ELA theory, methodology, evaluation, error analysis)
• PPT presentation for final review
• Viva Q&A preparation document (JPEG compression, DCT blocks, ELA interpretation, XGBoost, CASIA)
• Setup guide (environment, dataset download, running your own analyses)

• ELA only works on JPEGs: PNGs, screenshots and repeatedly recompressed images have no usable compression history, and the report states this plainly.
• Spliced regions saved at the same quality as the host photo can be invisible to ELA — the method detects inconsistencies, not manipulation itself.
• This is an analysis aid, NOT forensic evidence: it never substitutes for expert examination and must never be used to support legal claims.