Post-hoc Editing
Detects signs that a chart image was edited after creation, such as locally blurred areas, inconsistent color palettes, or compression artifacts that differ between the chart and modified regions.
Technical description
Screens for localized editing of a chart's own content with three signals that complement the whole-image forensics indicators. (1) Text/element-localized editing: recompresses the image at JPEG quality 95 and compares the Error Level Analysis (ELA) residual of each OCR text region (at least four needed) to the others; a region departing from the median by more than 2.0 and a robust z-score above 3.5 is an edited or spliced label, contributing min(2.5, outliers x 1.25). (2) Localized blur: a 16x16 block whose Laplacian variance is below one third of the global mean while all four neighbours are above it adds 1.5. (3) Color palette consistency: a block whose mean color exceeds max(mean + 2*std, 30) from the k-means dominant palette adds 1.0. Score capped at 5.0.
How it works
Layer 1 (deterministic). Recompresses the image and compares the ELA residual of each OCR text region to the others to flag edited or spliced labels; tiles the image to find an isolated blurred block; and compares block colors to the dominant palette to find pasted elements. Sums the three contributions, caps at 5.0, and reports the edited text region, blur, and color-outlier counts.
Why this matters
Editing a chart after the fact, changing a number, pasting a different bar, painting over a label, leaves localized traces even when the result looks seamless. Error Level Analysis exposes regions with a different compression history, a standard cue for splicing, copy-move, and retouching, and splicing is the most common, least-studied manipulation of scientific figures. G10 applies this where charts are actually edited: an edited number is a small region re-encoded into the figure, so its compression history diverges from the surrounding text, and a pasted element introduces a foreign color or a tell-tale blur.
Score thresholds
- 0-1
- Consistent compression across the chart's text, uniform sharpness, colors within the palette
- 2-3
- One signal: a text region with inconsistent compression, an isolated blurred block, or a foreign-color block
- 4-5
- Several signals together, consistent with post-hoc editing
Limitations
The text-editing signal needs at least four OCR-readable text regions, so sparse-text charts are screened only by blur and color. ELA depends on compression history: a losslessly saved or uniformly recompressed figure weakens the residual contrast, and a chart legitimately assembled from panels of different origin can show benign differences. Blur and color are directional cues, not proof. Whole-image Error Level Analysis, JPEG-ghost analysis, noise-residual consistency, and copy-move detection live in the dedicated image-forensics indicators (I1, I7, I3, I6), so G10 does not repeat them and stays on chart-element-localized editing.
References
- Krawetz N. (2007). A Picture's Worth: Digital Image Analysis and Forensics (Error Level Analysis). Black Hat USA Briefings
- Lin X, Tang W, Wang H, Liu Y, Ju Y, Wang S, Yu Z. (2024). Exposing image splicing traces in scientific publications via uncertainty-guided refinement. Patterns (Cell Press)
- Nagm AM, Moussa MM, Shoitan R, Ali A, Mashhour M, Salama AS, AbdulWakel HI. (2024). Detecting image manipulation with ELA-CNN integration: a powerful framework for authenticity verification. PeerJ Computer Science 10:e2205
- Nandi S, Natarajan P. (2026). Rescind: Countering Image Misconduct in Biomedical Publications with Vision-Language and State-Space Modeling. arXiv:2601.08040