01
Accepted media
An authorized image, video, audio track, or document enters a supported product workflow.
The accepted input depends on the product and published contract.Shared technology foundation
How AI deepfake detection technology analyzes supported media
Neural Defend products examine supported media for technical signals associated with AI generation, manipulation, or spoofing. This AI media detection process returns evidence for a product workflow or reviewer; it does not replace the decision that follows.
Plain-language definition
- Synthetic media
- Content generated wholly or partly by software, including generated images, video, voices, or text.
- Manipulated media
- Existing media changed in a way that can alter what a person sees, hears, or understands.
The technology looks for evidence in the media. It cannot infer a creator's intent from a file alone.
One layered view
Media moves through analysis. Decisions stay outside it.
This conceptual diagram explains the shared pattern at a useful level. It is not a disclosure of proprietary model architecture, preprocessing, routing, or signal-combination rules.
Conceptual system viewInput → evidence → customer action
- 02
Input checks and standardization
The system checks whether the media can be analyzed and prepares a consistent analysis input.
Exact preprocessing and routing details are not public. - 03
Detection signals
Relevant models examine technical patterns in the visual, audio, temporal, or content data.
The exact model inventory and internal architecture are not public. - 04
Risk and evidence
Supported signals contribute to result fields such as status, risk, and linked evidence.
This is not an independent statement of fact about the media. - 05
Customer workflow
Customer policy or an authorized reviewer decides what happens next.
Thresholds, escalation, and final action remain customer-owned.
Signals and decisions
A signal is an observation, not a verdict.
A detection signal is a measurable response to a technical pattern in supported media. Several signals can matter because generation, editing, capture, compression, and delivery affect files in different ways. Signals can also conflict or be unavailable.
What the system supplies
Evidence for the next step
- Supported status, risk, or modality fields
- Signal-level or localized findings where available
- Explicit rejection and error states
- Model and version context in supported review workflows
What remains outside the result
The decision and its consequences
- Whether a person, payment, claim, or account is approved
- Who created the media or what they intended
- Whether the depicted event or attached claim is true
- Which threshold, fallback, appeal, or review path applies
Supported AI media detection
AI-generated content detection is limited by product scope.
Here, AI-generated media detection and synthetic media detection refer only to inputs documented for each product. They do not imply a public text-detection endpoint or universal content coverage.
Images
Public NeuroVerify contractThe published image endpoint returns status and image risk fields for supported requests.
PublishedVideo with audio track
Public NeuroVerify contractThe published video response contains separate video and audio-track result fields.
PublishedStandalone audio
Product-specific capabilityNeuroDetect reviews audio, and NeuroMeeting is positioned for live audio and video analysis. The public NeuroVerify contract does not publish a standalone audio endpoint.
No standalone API endpointDocuments
NeuroDetect review capabilityNeuroDetect supports document investigation. The public NeuroVerify contract does not publish a standalone document endpoint.
Not in public API contractC2PA validation
Unconfirmed product featureC2PA is useful for provenance, but the public NeuroVerify contract does not confirm a C2PA validation endpoint.
Do not assume supportProduct-family status
Beta and Coming Soon describe availability. They should not be read as proof of a production deployment or confirmed public interface.
NeuroVerifyLive
API-based image and video authenticity results under the published contract.
NeuroDetectLiveAnalyst-led review of image, audio, video, and document evidence.
NeuroCallBetaLive voice analysis and customer-controlled signal routing for supported live-call workflows.
NeuroMeetingComing SoonPlanned participant-level audio and video signals for live-meeting workflows.
Model and system evaluation
Test the conditions that can change the result.
A useful evaluation names the task, media, attacks, model version, operating conditions, metrics, and decision costs. Without those details, a performance number cannot show whether the system fits a real workflow.
Read the evaluation approach01
Representative attacks
Test genuine, synthetic, manipulated, recompressed, recaptured, low-quality, and unsupported examples that resemble the intended workflow.
02
Unseen generators
Keep generator families or attack methods outside the development set, then measure how the evaluated version handles unfamiliar examples.
03
Errors and uncertainty
Record false positives, false negatives, rejected inputs, service errors, and inconclusive results rather than collapsing them into one accuracy figure.
04
Thresholds and review
Choose thresholds against the cost of each error, available fallback checks, and the number of cases a reviewer can examine.
05
Latency under real conditions
Measure response time with the expected formats, file sizes, media quality, concurrency, deployment path, retries, and timeouts.
No universal performance value is asserted here.
Accuracy, error rates, latency, and review volume depend on the evaluated product version, dataset, threshold, media quality, traffic, hardware, and deployment conditions.
Complementary approaches
Detection answers one part of an evidence question.
Explainability, provenance, and contextual verification examine different questions. Strong review keeps their findings separate before bringing them together.
01
Explainability
Where should a reviewer look?
A supported view can point to a region, frame, interval, page, or signal. It does not independently establish authorship, intent, source identity, or event truth.
02
Provenance
What origin or edit-history claims can be validated?
C2PA can validate signed, tamper-evident assertions under a trust configuration. A valid assertion does not establish that the depicted event or attached caption is true.
03
Context
Does the claim fit the surrounding evidence?
A reviewer compares the media with its source, timeline, earlier versions, related records, and independent corroboration. Detection cannot supply this context by itself.
Privacy and security design
Define the operating boundary before deployment.
Product selection does not answer every privacy or security question. Teams should document the specific service, deployment, data path, controls, and responsibilities they intend to use.
Questions to answer
- 01
Which media types are accepted, rejected, or retained for the selected product?
- 02
Where is media processed and stored for the proposed deployment?
- 03
Who can submit evidence, inspect results, change thresholds, or export records?
- 04
What deletion, logging, key-management, backup, and incident processes apply?
- 05
How are model, software, policy, and configuration changes recorded and reviewed?
- 06
Which responsibilities remain with the customer, hosting provider, and Neural Defend?
Operating limitations
Quality changes the available evidence
Compression, cropping, re-encoding, noise, short duration, poor lighting, and missing metadata can weaken or alter useful signals.
Unknown attacks remain possible
A low-risk result applies to the evaluated model, supported input, and current conditions. It does not prove that media is authentic.
A high-risk result is not motive or identity evidence
Detection can support a manipulation review. It cannot establish who created the media, why they created it, or whether every claim around it is false.
Some requests should not produce a score
Unsupported, damaged, poor-quality, unavailable, or contradictory inputs need an explicit rejection, error, or review path.
Continue the evaluation
Move from shared concepts to a specific product boundary.
Use the canonical evaluation and deployment guidance, then confirm current product behavior against the intended workflow.
Evaluation
Plan a representative technical evaluation.
Define the protected decision, authorized media, attack coverage, error costs, thresholds, and acceptance checks.
Deployment
Map the operating boundary.
Review data paths, access, retention, connectivity, updates, support, and customer responsibilities.
Products
Choose the product workflow.
Compare the live, Beta, and planned Neural Defend product surfaces.
Research
Continue to published material.
Read current articles and research before evaluation.
Technical evaluation
Bring the media, conditions, and decision that matter.
A scoped evaluation can confirm current product coverage, response handling, thresholds, operating limits, deployment responsibilities, and review criteria without relying on a generic benchmark.