An image file does not inherently declare whether its pixel values came from a physical imaging sensor or a synthetic generation process. SPOT—Sensor Pattern Origin Testing—examines characteristics of image formation to assess those competing origin propositions. This is the particular problem that connects RHEM Labs’ work to searches for AI-generated image detection, synthetic-image detection and image authentication.
The search terms describe overlapping questions. Deepfake detection can concern face replacement, manipulated video, synthetic audio or fully generated imagery. Image provenance concerns a record of where a file came from and how it changed. SPOT addresses a narrower image-origin question: what support do measurable image characteristics give to physical camera capture rather than synthetic generation? It does not purport to detect every kind of deepfake or to authenticate the truth of the depicted scene.
Visual appearance alone does not establish origin. Metadata and Content Credentials can provide useful provenance evidence, but may be absent or interrupted by subsequent processing. Sensor-related measurements provide another line of enquiry. SPOT is RHEM Labs’ forensic media-origin assessment capability: its reporting expresses the relative support of observations for defined hypotheses rather than converting a score into an unqualified “real” or “fake” label.
An implemented prototype exists, supported by published experimental research and application to questioned real-world imagery. The next stage is representative operational evaluation: establishing how the capability generalises, integrates and creates value in the environments where it would be used.
Evidence left by image formation
A conventional digital photograph begins with light passing through optics and interacting with a semiconductor image sensor. Conversion into electrical signals is affected by sensor behaviour, electronic noise and the camera’s image-processing pipeline. These processes can leave measurable characteristics in the resulting image.
A wholly synthetic image has a different formation history. It can contain noise and statistical structure, but these arise through a different process. SPOT builds on research into physical image formation, sensor behaviour and sensor pattern noise to investigate whether those differences support forensic assessment of origin.
Physical image formation: sensor cross-section from Richard Matthews’ 2019 doctoral thesis, Figure 5.2, p. 61. This is foundational sensor research, not an image of the SPOT prototype. Read the thesis.
The peer-reviewed method by Dr Richard Matthews converts imagery to grayscale and suppresses scene content to extract image noise residues. It measures variance, skewness and kurtosis: characteristics describing the residue’s spread, asymmetry and tendency towards extreme values. Camera and synthetic-image populations are statistically modelled, then a questioned image’s measurements are evaluated against both. This is distinct from matching a questioned image to the fingerprint of a particular camera.
A likelihood ratio expresses the relative support those measurements provide for the competing origin hypotheses under the models. It is evidence about origin, not a probability that an image is “real”.
Strong discrimination in a defined experiment
The research published in the Australian Journal of Forensic Sciences compared Dresden Image Database camera imagery with DALL·E 3 synthetic imagery, using separate training and test samples.
Across 100 random runs, the published Results section reports mean ROC AUC of 0.998, accuracy of 0.985, precision of 0.987 and recall of 0.983. A representative 3,000-image test correctly classified 1,486 of 1,500 camera images and 1,489 of 1,500 synthetic images.
One published test example, with 1,500 images of each origin. Taking camera origin as the positive class: 1,486 true positives, 14 false negatives, 11 false positives and 1,489 true negatives. These counts are separate from the means across 100 runs. Source: Results section, p. S70.
These are strong empirical results from a defined experimental population. They establish that the scientific hypothesis has been implemented and produced substantial discrimination under the tested conditions. They do not establish universal operational accuracy or performance across every camera, generator and processing history.
Explore the technical evidence and experimental evaluation or read the publication record and publisher link. The forensic perspective on AI-generated image detection places this approach alongside provenance and other detection methods.
Questions adjacent to SPOT
Readers asking how to tell whether an image was AI-generated need to distinguish visual clues, metadata, signed provenance and measured origin evidence. The C2PA and Content Credentials analysis examines the particular problem of missing or broken provenance. Neither path turns absent metadata into a synthetic-origin finding.
The broader phrase deepfake detection includes audio, video and local manipulation questions outside SPOT’s present still-image scope. For decisions that depend on a detector output, error rates, thresholds and base rates need separate treatment. The articles on AI-generated images as evidence and investigative workflow connect those measurements to acquisition and reporting.
Applied examination: why evidential strength matters
In a recent applied examination, RHEM Labs was asked to assess questioned imagery from a real-world matter. The laboratory independently located original source imagery from the best available source, preserving provenance as far as practicable and reducing contamination through subsequent processing.
Two images were examined. The first produced a likelihood ratio of approximately 4.27 billion to one in favour of physical camera origin over the competing synthetic-origin proposition: very strong support under the applied model.
The questioned image produced approximately 5:1 in the same direction. Although the observation favoured physical camera origin, the strength of that support was weak and was not considered sufficient to support a conclusion of physical camera origin.
Both results favoured the same proposition numerically; their evidential significance differed markedly. SPOT did not determine that either image was “real” or “fake”. It assessed the strength of the origin evidence. Results may offer strong, moderate or weak support, little useful discrimination, or fall outside conditions supporting a useful conclusion.
Origin also remains distinct from truth. A camera can photograph a screen, print or synthetic image. Support for physical capture does not establish that an event occurred as represented, nor resolve separate questions of partial manipulation, editing, provenance or authorship.
The complete examination, including locating suitable imagery, examination and interpretation, took approximately ten minutes. SPOT processing and result generation took seconds. This documents applied use beyond the original experimental dataset, without establishing population-wide operational validation or a standard turnaround for future matters.
From origin evidence to forensic advantage
Rapid origin information creates the possibility of using forensic evidence early in a workflow. Measured image characteristics become an assessment of competing propositions; that assessment provides confidence-based information that can reduce uncertainty and guide triage, further examination or a decision.
For Defence and intelligence, this could strengthen information resilience and analytical confidence by helping analysts assess imagery before relying on it. Detecting questionable origin evidence early could reduce the effectiveness, or increase the cost, of synthetic-media-enabled deception. That potential capability effect requires operational evaluation.
For law enforcement and investigators, the contribution is evidence assurance: additional scientific information about questioned imagery and a framework for communicating its strength. Regulators and financial institutions face the same need where imagery supports claims, fraud, compliance, integrity or financial-crime enquiries.
Across these settings, SPOT could serve as a triage and evidence-assurance capability within a larger analytical system, helping direct material to specialist or more computationally expensive examination alongside other evidence and professional judgement.
A capability grounded in forensic science
RHEM Labs is an independent Australian forensic science laboratory whose commercial foundation is forensic examination, expert reporting and expert evidence. Research and engineering address problems exposed through that practice.
SPOT’s foundation extends to Dr Richard Matthews’ doctoral research into physical digital-image formation and sensor behaviour. That foundation has been extended through peer-reviewed synthetic-media research and translated into an executable prototype, controlled experimental evaluation and applied examination.
The expertise, research capability and intellectual property needed to mature SPOT remain Australian-held. This provides a sovereign foundation for further development and integration.
Explore the underlying research and the laboratory’s expertise.
From experimental capability to operational use
Representative operational evaluation should establish performance against the imagery a user actually encounters: unfamiliar cameras and generators, compression, resizing, screenshots, physical recapture, mixed provenance and partial editing. It must identify failure and inconclusive conditions and test how evidential strength and uncertainty should be communicated.
Evaluation must also establish where SPOT belongs in the analytical pipeline, which material proceeds to deeper examination, and what interfaces and deployment architecture are required. The operational test is whether earlier origin information improves triage, prioritisation, analytical confidence or another defined outcome.
RHEM Labs is engaging operational users, integration partners and independent evaluators to undertake representative evaluation and capability integration. Operational users can contribute use cases, representative datasets and reporting requirements. Technical partners can define interfaces with existing systems. Independent evaluators and research partners can strengthen experimental design and representative testing.
The capability exists. The next programme concerns generalisation, integration and capability effect. Contact RHEM Labs to discuss an origin-assessment problem, representative evaluation, integration requirements or a scoped capability-development programme.
