Sensor Pattern Noise

SPN is caused by silicon wafer imperfections during sensor fabrication, creating pixel-to-pixel variations in sensitivity. This pattern is:

• Invariant to temperature and humidity within normal operating ranges
• Persistent across the entire lifetime of the sensor
• Unaffected by JPEG compression, lens changes, or firmware updates

This stability makes it superior to metadata for source attribution, as the fingerprint cannot be stripped or altered by file format conversion.

Unlike additive thermal noise, SPN follows a multiplicative model where the noise amplitude scales with the incident light intensity. The dominant component, Photo Response Non-Uniformity (PRNU), is extracted using:

• A denoising filter to separate the noise residual from the image content
• Maximum likelihood estimation across multiple reference images to suppress scene content

The PRNU factor is estimated as the difference between the original image and its denoised version, normalized by the clean image.

SPN manifests as a pixel-level, high-frequency pattern that is invisible to the human eye but statistically robust:

• Pixel non-uniformity typically ranges from 1% to 2% of the signal
• The pattern is spatially uncorrelated, resembling white Gaussian noise
• Contains energy primarily in the high-frequency domain, making it separable from low-frequency scene content

Forensic analysts exploit this frequency separation using wavelet-based denoising to isolate the SPN from semantic image structures.

Source identification is performed by computing the normalized cross-correlation between the extracted noise residual and a reference SPN pattern:

• Peak to Correlation Energy (PCE) ratio is the standard detection statistic
• A PCE value above 60 typically indicates a definitive match
• The correlation is computed in the spatial domain after zero-meaning both signals

This statistical test is robust to cropping, resizing, and moderate geometric transformations when re-synchronization is applied.

Beyond source identification, SPN enables pixel-level forgery localization by detecting regions where the expected sensor fingerprint is absent:

• The image is divided into sliding blocks (typically 64×64 or 128×128 pixels)
• Each block's noise residual is correlated with the reference SPN
• Absent or weak correlation indicates a region originating from a different sensor or synthetically generated

This sliding window approach produces a heatmap identifying spliced or inpainted areas.

SPN survives most standard image processing operations that destroy metadata-based attribution:

• JPEG compression up to quality factor 70 preserves detectable correlation
• Gamma correction and white balance adjustments do not eliminate the multiplicative pattern
• Resizing requires re-synchronization but the fingerprint remains extractable

However, strong geometric transformations (rotation, perspective warping) and aggressive denoising can attenuate the signal, requiring advanced synchronization techniques.

The dominant source camera identification method that directly leverages sensor pattern noise. PRNU is caused by pixel non-uniformity—microscopic variations in the sensitivity of individual photodiodes on a silicon sensor.

• Stochastic fingerprint: Acts as a unique, deterministic biometric for the sensor
• Multiplicative noise: Scales with illumination, extracted via a denoising filter
• Robustness: Survives re-compression, resizing, and gamma correction

The correlation between an image's noise residual and a camera's reference PRNU pattern provides a statistical match score for device attribution.

A deep learning-based camera model fingerprint that moves beyond raw PRNU to capture the relationship between noise residuals and image semantics. A Siamese network extracts a noiseprint—a 2D map that encodes local noise characteristics.

• Tampering localization: Deviations in the noiseprint reveal spliced regions
• Model-level attribution: Identifies the camera model without a reference device
• Content-aware: Adapts to texture, edges, and homogeneous areas

Unlike PRNU, noiseprint does not require a reference image from the suspect device, making it effective for blind forgery detection.

Transforms an image from the spatial domain into its frequency representation using the Discrete Fourier Transform (DFT) or Discrete Cosine Transform (DCT). Sensor pattern noise manifests as high-frequency components that are invisible to the naked eye.

• Periodic artifacts: Grid-like patterns from CFA interpolation become visible
• Noise spectrum: Authentic sensor noise has a characteristic power-law distribution
• Anomaly detection: Synthetic images often lack natural high-frequency noise decay

Frequency analysis is a critical preprocessing step for isolating the noise residual before PRNU extraction.

The forensic task of determining the make and model of the source camera from an image alone. This technique analyzes a composite of proprietary in-camera processing traces.

• Sensor pattern noise: PRNU provides the strongest individual signal
• CFA interpolation: Each manufacturer uses proprietary demosaicing algorithms
• JPEG quantization tables: Brand-specific compression fingerprints
• Lens distortion: Chromatic aberration and vignetting patterns

Camera model identification narrows the search space before performing individual device matching via PRNU correlation.

The forensic process of identifying boundaries where a donor region from one image has been inserted into a host image. Sensor pattern noise is a powerful cue because the spliced region carries the PRNU of a different sensor.

• Noise inconsistency: Local SNR deviates from the host image's noise profile
• Block-level PRNU correlation: Sliding window analysis reveals mismatched fingerprints
• Edge discontinuity: High-pass filtering exposes abrupt noise transitions

When the donor camera is unknown, statistical anomalies in the noise residual serve as the primary tampering indicator.

The forensic task of generating a pixel-level binary mask that precisely identifies manipulated regions within an image. Unlike global authenticity classification, localization requires spatial precision.

• PRNU prediction error: A camera's reference pattern is subtracted; anomalous residuals indicate forgery
• Sliding window correlation: Computes local PRNU similarity across the image grid
• Bayesian MAP estimation: Fuses noise, CFA, and compression cues into a probability map

The output is a heatmap or binary mask highlighting every pixel inconsistent with the purported source device's sensor pattern noise.