Copy-Move and Splicing Detection in Images

Every digital image is a statistical object. Its pixel values, their spatial correlations, their frequency-domain distribution after compression, and the pattern of sensor noise they carry all reflect the specific physics of the capture device and the specific sequence of processing operations applied. Forgery breaks one or more of these statistical regularities.

For copy-move, the break is internal consistency. A genuine photograph taken by a single camera has one noise level across the frame, one demosaicing pattern, one focus blur distribution, and one JPEG compression grid.

When a patch is copied and pasted, the pasted area and the original area have statistically identical texture and noise, a coincidence that is essentially impossible in a genuine image. If the forger applies geometric transformations such as rotation or scaling to disguise the copied region, the transformation itself leaves traces: interpolation artifacts, resampling periodicity, and a changed local noise spectrum.

For splicing, the break is cross-region consistency. A composited image contains content from cameras with different noise characteristics, different colour matrices, different JPEG quality settings, and potentially different acquisition conditions (lighting angle, colour temperature).

The boundary between the spliced region and the background may show an abrupt change in noise variance, a misalignment of the JPEG discrete cosine transform (DCT) blocking grid, or a sharp transition in the sensor's photo-response non-uniformity (PRNU) pattern. Detection methods target each of these inconsistency types.

Block-matching is the classical approach to copy-move detection, introduced in the work of Fridrich et al. in 2003. The algorithm divides the image into overlapping blocks of fixed size, typically 16 by 16 pixels, and represents each block as a compact feature vector.

DCT coefficient vectors and PCA-projected pixel vectors are the most common representations. The full set of feature vectors is then sorted lexicographically, and adjacent entries in the sorted list are compared. Pairs of blocks whose feature vectors fall within a small distance threshold are flagged as candidate copies.

From the candidate pairs, the algorithm computes offset vectors: each pair has a spatial offset equal to the vector from one block's position to the other's. Genuine block pairs that arise from chance similarity are scattered across many offset values. Cloned regions produce clusters of pairs sharing a consistent offset, because every block in the source region and its copy is displaced by the same translation. Post-processing uses this clustering to suppress false matches and produce a binary localisation map.

Block-matching has two well-known failure modes. First, geometric transformation: if the forger rotates, scales, or shears the copied region before pasting, the feature vectors of corresponding blocks no longer match under direct comparison.

Extended variants using rotation-invariant features (Zernike moments, log-polar transforms) address this at a computational cost. Second, JPEG re-compression: saving the forged image as JPEG alters pixel values throughout, including in the copied region. The DCT rounding introduced by compression can push block pairs that were identical before save below threshold. Using a looser similarity threshold recovers some of these pairs but increases false positives.

Keypoint-based methods apply local feature descriptors developed originally for object recognition and image stitching to the problem of finding self-similar regions within a single image. SIFT (Scale-Invariant Feature Transform) detects interest points at multiple image scales and computes 128-dimensional gradient-based descriptors that are invariant to scale, rotation, and moderate illumination change. SURF and ORB (Oriented FAST and Rotated BRIEF) are faster variants with similar intra-image matching capability.

To detect copy-move with SIFT, the examiner extracts keypoints from the entire image and matches descriptors within the same image. In a genuine photograph, matched pairs arise from natural textures (brick walls, fabric) that have repeated structure, but the geometric relationship between matched pairs is inconsistent: they point in many directions and distances.

In a copy-move forgery, the matched pairs involving the cloned region produce a geometrically consistent cluster, because all the copies are displaced by the same affine transformation from their source. Grouping matches by homography consistency, for example using RANSAC (Random Sample Consensus), isolates the forgery cluster from background matches.

Keypoint methods handle rotated and scaled copies well. SIFT descriptors are computed relative to the dominant local gradient orientation, so a region copied and rotated 45 degrees still produces descriptors that match the source region's descriptors.

This makes keypoint methods considerably more resilient than basic DCT block-matching against geometric post-processing. Their limitation is density: keypoints concentrate on edges and textures, so a copied region that falls in a smooth, featureless area (sky, a painted wall) generates few or no keypoints and may be missed entirely. Block-matching covers featureless regions but fails on geometric transformation; combining both methods mitigates the blind spots of each.

From around 2015, convolutional neural networks began outperforming hand-crafted methods on standard forgery benchmarks. The early approach trained a CNN end-to-end on image patches labelled authentic or forged.

Subsequent architectures improved generalisation by feeding the network a noise residual (the image minus a denoised version, computed with a fixed filter such as a constrained convolutional layer) rather than raw pixels. Noise residuals amplify manipulation artifacts while suppressing semantic image content, making the network sensitive to the statistical traces of editing rather than to object appearance.

More recent models use dual-stream or frequency-domain inputs. One stream processes the RGB image; the other processes DCT coefficients or a high-pass filtered version. Fusion layers combine both streams to classify each image region. Transformer-based architectures have been applied to detect long-range inconsistencies that local convolutional filters miss. In controlled benchmark evaluations, these models achieve pixel-level localisation AUC above 0.95 on standard datasets such as COVERAGE, CASIA, and Columbia Uncompressed.

Splicing detection with deep learning follows the same structure but targets boundary artifacts rather than internal self-similarity. Networks learn to flag transitions in noise texture, lighting consistency, and colour statistics at region boundaries. The approach is sensitive to sophisticated compositing that blends these statistics at the boundary, and practitioners must be cautious about false negatives when the forger has used a feathered edge, gradient blend, or colour-matching tool.

A casework examination begins with the questioned image in its originally received form. Any processing applied to the evidence file must be documented and reversible. The examiner first runs file-level checks: EXIF metadata, JPEG quantisation table, thumbnail-main discrepancy, and error-level analysis. These preliminary checks establish the image's processing history and may reveal obvious manipulation before algorithm-level detection begins.

Detection is then applied in layers. Block-matching is fast and covers featureless regions; keypoint matching covers geometric transformations; noise residual analysis targets splicing boundaries; PRNU analysis (where a reference noise pattern from the claimed source camera is available) tests device attribution.

Each method produces a localisation map. Convergence across maps strengthens the finding: if three independent methods flag the same region, the probability of all three producing a false positive simultaneously is very low. Divergence between maps indicates either a method-specific false positive or a complex forgery combining multiple techniques.

The examination report documents: the received file hash (SHA-256), the software and algorithm versions used, the parameter settings (block size, threshold, keypoint detector version), the output localisation maps, and a plain-language interpretation. Where the examiner concludes that manipulation is detected, the report quantifies confidence and lists alternative explanations considered and ruled out. Where results are inconclusive, this must be stated explicitly.

Image authentication findings are admitted as expert evidence under jurisdiction-specific rules. In the United States, the Daubert standard requires that expert testimony be grounded in a testable, peer-reviewed methodology with a known or estimable error rate. In England and Wales, the Law Commission's report on expert evidence and the Criminal Practice Directions require disclosure of the basis and methodology.

In India, the Bharatiya Sakshya Adhiniyam 2023 (BSA 2023, which replaced the Indian Evidence Act 1872) governs the admissibility of electronic evidence and requires that digital records be certified as authentic by a person responsible for the device or process. Across the European Union, the eIDAS regulation and national procedural codes create comparable admissibility frameworks for electronic evidence.

Common challenges during testimony include: the opposing expert claiming the detection software is a black box with unknown false-positive rates; the suggestion that JPEG compression artefacts account for the flagged region; and the argument that the detection method has not been validated on images of this type. The examiner answers these by citing peer-reviewed validation studies, providing the software's published false-positive benchmarks, and describing the convergence of multiple independent methods in the specific examination.

International cases, such as those before the International Criminal Court (ICC) or the European Court of Human Rights, require the examiner to explain the methodology without assuming familiarity with any particular national evidentiary framework. The ICC has received forensic image authentication evidence in several proceedings, and the practice direction is to present the method's scientific basis, the output, and its limitations in terms that a tribunal of any legal tradition can evaluate.