Splicing Detection

Every camera sensor introduces a unique, stochastic noise pattern during image capture. When a region from a donor image is spliced into a host, the local noise characteristics—such as variance, distribution, and correlation—will differ from the surrounding authentic areas.

• Local Noise Estimation: Algorithms estimate the noise residual by denoising the image and subtracting the clean version. Inconsistencies in the residual map highlight potential splices.
• Sensor Pattern Noise (SPN): The photo response non-uniformity (PRNU) acts as a camera fingerprint. A spliced region will lack the host camera's PRNU or contain a foreign device's pattern.
• Statistical Moments: Higher-order statistics (skewness, kurtosis) of noise residuals are computed in sliding windows to detect abrupt transitions at the splice boundary.

Most digital images are stored in the JPEG format. Splicing operations typically require decompressing both the host and donor images, performing the manipulation, and re-compressing the result. This process leaves distinct forensic traces.

• Double JPEG Compression: When a spliced region has a different compression history than the host, the quantization tables will be inconsistent. Detecting the presence of a primary and secondary quantization grid reveals the edit.
• Blocking Artifact Grid Misalignment: JPEG compresses images in 8x8 pixel blocks. A spliced region often exhibits a shifted or misaligned block grid relative to the host, visible in the spatial or DCT domain.
• Quantization Error Analysis: The distribution of DCT coefficient errors after re-compression will be anomalous at the splice boundary, as the donor region's coefficients are quantized with a different step size.

Digital cameras use a Color Filter Array (typically a Bayer pattern) over the sensor, so each pixel captures only one color. The missing two colors are estimated via demosaicing interpolation. This process creates predictable statistical correlations between neighboring pixels.

• Interpolation Kernel Estimation: A linear model is used to predict a pixel's value from its neighbors. The prediction error will be low in authentic regions but high in spliced areas where the donor image's interpolation pattern is different or absent.
• Periodic Correlation Patterns: The demosaicing process introduces periodic artifacts in the frequency domain. A spliced region will exhibit a different spectral signature, detectable via Fourier analysis.
• Variance Map Inconsistency: The variance of the prediction error is mapped across the image. Abrupt changes in this variance map indicate a transition between different interpolation algorithms or the absence of one.

Splicing creates a physical boundary between the host and donor regions. Even with careful blending, subtle discontinuities in edge profiles and gradient fields often persist at the splice boundary.

• Gradient Vector Flow (GVF): The gradient field is computed across the image. A splice boundary introduces a high-energy, closed-contour anomaly in the gradient magnitude that does not correspond to a natural object edge.
• Laplacian Edge Detection: The second-order derivative (Laplacian) is highly sensitive to abrupt intensity transitions. A thin, unnatural ridge in the Laplacian response often traces the exact splice contour.
• Edge Profile Blur Inconsistency: Authentic edges have a consistent blur profile determined by the camera's optics and sensor. A spliced region's edges will have a different blur point spread function (PSF), detectable by measuring edge width and slope.

A physically plausible scene has a consistent three-dimensional light environment. When a donor object is spliced from a scene with different illumination, the lighting cues on the object will conflict with the host environment.

• Light Direction Estimation: The angle of incident light is estimated from shading gradients on an object's surface. A spliced object will often have a light direction vector that differs from other objects in the scene.
• Shadow Analysis: Cast shadows must be geometrically consistent with the light source position and occluding object shape. A missing, misaligned, or incorrectly oriented shadow is a definitive sign of splicing.
• Specular Highlight Matching: The position and shape of specular highlights on the eyes, skin, or reflective surfaces must be consistent. A mismatch in highlight location between two eyes or between a face and the environment indicates compositing.
• Color Constancy & White Balance: Different cameras apply different white balance corrections. A spliced region may exhibit a subtle color temperature shift relative to the host, detectable by analyzing the illuminant color matrix.

Modern forensic systems use convolutional neural networks (CNNs) trained end-to-end to produce a pixel-level tampering localization mask, bypassing handcrafted feature engineering.

• Noiseprint Extraction: A Siamese network architecture extracts a camera model fingerprint that captures the relationship between noise residuals and image semantics. Anomalies in this fingerprint directly localize spliced regions.
• Constrained Convolutional Layers: Specialized layers suppress the image's semantic content while amplifying manipulation artifacts, forcing the network to learn forensic features rather than scene content.
• Multi-Scale Feature Fusion: Architectures like U-Net or Feature Pyramid Networks combine low-level texture cues (noise, edges) with high-level semantic context to produce high-resolution, artifact-free localization maps.
• Adversarial Robustness: Forensic classifiers are trained with adversarial examples to resist attempts to fool them with anti-forensic perturbations designed to hide splicing traces.