Spatial Rich Model (SRM)

SRM projects an image through a diverse bank of linear and non-linear high-pass filters to compute noise residuals. These residuals suppress the image's semantic content and isolate the subtle, local pixel dependencies introduced by the sensor or manipulation.

• Linear Filters: Capture first-order and second-order differences, revealing edge artifacts.
• Non-Linear Filters: Include 'minmax' and 'spam' filters that capture higher-order statistical deviations invisible to linear analysis.
• Diversity: The model uses dozens of distinct filters to ensure no single type of manipulation artifact is missed.

After residual computation, SRM quantizes the continuous residual values and computes fourth-order co-occurrence matrices. These matrices model the joint probability distribution of neighboring residual samples, capturing the intricate statistical texture of the noise.

• Truncation: Residuals are truncated to a small range (e.g., [-T, T]) to control dimensionality and suppress outlier noise.
• Symmetry Aggregation: Horizontal, vertical, diagonal, and anti-diagonal co-occurrences are merged using sign and directional symmetries to reduce the feature space while preserving discriminative power.
• Output: A single, massive feature vector representing the global statistical fingerprint of the image's noise.

Unlike forensic methods targeting a specific manipulation (e.g., double JPEG compression), SRM is a universal blind detector. It does not require prior knowledge of the manipulation type.

• Rich Model Concept: The 'richness' refers to the massive, diverse feature set that over-represents the image, ensuring that any manipulation's footprint is captured in some subspace of the model.
• Generalization: Trained on a broad mix of manipulations, an SRM-based classifier can detect previously unseen inpainting, splicing, or steganographic algorithms by identifying a break in the natural statistical structure of the sensor noise.

The 34,671-dimensional SRM feature vector is fed into an ensemble classifier, typically a set of Fisher Linear Discriminants (FLDs) trained on random subspaces of the feature space.

• Dimensionality Management: Training on random subspaces prevents overfitting to the high-dimensional, sparse feature set.
• Fusion Rule: The final authenticity decision is made by a majority voting mechanism across the ensemble, providing a robust and computationally feasible classification.
• Scalability: This architecture allows the model to scale to the massive feature dimensionality required for universal detection without prohibitive training costs.

SRM was originally designed for steganalysis (detecting covert hidden messages) but is equally effective for image manipulation detection.

• Steganalysis: Detects the subtle disruption of natural pixel dependencies caused by embedding a payload in the least significant bits.
• Manipulation Detection: Identifies the boundary where a spliced object introduces a foreign noise distribution, or where inpainting creates statistically anomalous smooth regions.
• Unifying Principle: Both tasks rely on detecting a localized violation of the camera's intrinsic sensor pattern noise and processing fingerprint.

SRM operates entirely in the spatial domain, analyzing pixel values directly without transforming the image into a frequency representation like DCT or wavelet domains.

• Complementary Analysis: While frequency domain methods (e.g., double JPEG detection) target compression artifacts, SRM targets the fundamental pixel correlations that exist before and after compression.
• Robustness: This spatial focus makes SRM effective even on high-quality, minimally compressed images where JPEG ghosting artifacts are absent.
• Feature Submodels: The 'SRMQ1' variant includes quantization steps that provide some implicit robustness to JPEG compression while remaining a spatial-domain feature.

The foundational preprocessing step for SRM, where an image's denoised version is subtracted from the original to isolate the stego noise component. This suppresses scene content and amplifies the subtle artifacts introduced by manipulation. Common filters include wavelet-based denoisers and pixel prediction error filters that model local linear dependencies.

A statistical tool that captures the joint probability distribution of neighboring residual values. Instead of analyzing individual pixels, it models the higher-order relationships between adjacent samples in the noise residual. SRM uses fourth-order co-occurrence matrices, which track the frequency of specific 4-pixel patterns, dramatically increasing the feature space to capture complex textural anomalies.

The machine learning backbone typically paired with SRM features. An ensemble of base learners—often Fisher Linear Discriminants (FLD) trained on random subspaces of the high-dimensional feature set—is fused by majority voting. This architecture is specifically designed to handle the small training sample size, high dimensionality problem inherent in steganalysis and forgery detection, providing robust generalization.

A dimensionality reduction technique applied to the raw SRM feature vector. Instead of using the full 34,671-dimensional feature set, random projections are used to map the co-occurrence bins into a lower-dimensional space. This preserves the relative distances between feature vectors while making the subsequent classifier training computationally tractable without significant loss of forensic discriminability.

While SRM was originally designed for steganalysis—detecting hidden messages embedded by steganographic algorithms—its principles transfer directly to image forgery detection. Both tasks rely on identifying subtle statistical anomalies in the noise domain. The key distinction: steganalysis looks for embedding artifacts, while forgery detection uses SRM to localize tampering boundaries where noise patterns are inconsistent.

A critical consideration when applying SRM to real-world images. The model must account for the 8x8 block grid introduced by JPEG compression. SRM incorporates features computed directly from JPEG quantized DCT coefficients, capturing intra-block and inter-block dependencies. This phase-aware variant prevents the forensic classifier from confusing legitimate compression artifacts with manipulation traces.