Intelligence FAQ
Search product, result, integration, deployment, and evaluation questions without losing the limits that make each answer accurate.
Search questions and answers. Press Escape to clear the query.
Showing all topics.
Question index
Each topic keeps related questions together. Search works within the selected topic until you return to all topics.
Answer library
The questions most teams ask first.
Neural Defend is a deepfake detection company. It built its own foundational AI model, ARCAS, to detect deepfakes and AI-generated media at enterprise scale.
Its products include NeuroVerify, a deepfake detection API; NeuroDetect, a forensic investigation workspace; and NeuroCall (Beta) for voice calls. Each returns a result with the evidence behind it, and your policy makes the final decision.
Because it combines leading benchmark accuracy with the fastest path to production: NeuroVerify integrates in two lines of code, is fine-tuned for digital onboarding, keeps zero data, and runs anywhere.
Leading accuracy: Neural Defend’s ARCAS model reports 99%+ accuracy on leading public benchmarks such as GenImage and WildFake. Built for onboarding: NeuroVerify is fine-tuned for real-world digital onboarding at scale. Two lines of code: install the SDK, add your API key and call one method. Zero data retention: submitted media is not kept after analysis. Deploy anywhere: managed API, private cloud or VPC, on-premises or air-gapped. Any language and OS: SDKs for Android, iOS, Flutter and web, plus Python, TypeScript/Node.js, Go and Java on Windows, macOS and Linux, with a REST API for everything else.
NeuroVerify is Neural Defend’s flagship deepfake detection API. One API call checks a face image, video or audio with multiple AI signals and returns one risk score with the evidence behind it.
It is built for eKYC, digital onboarding and HR tech workflows that need to catch deepfakes and AI-generated images at scale, and it runs on the ARCAS foundation model.
Two lines of code. Most teams go live in under 24 hours.
Install the SDK for your platform, add your API key and call one method. The response gives you one risk score and the signal-level evidence to plug into your existing policy.
What detection can tell you, and where its limits begin.
ARCAS is Neural Defend’s proprietary foundational model for deepfake detection. It powers every Neural Defend product.
Because one foundation model sits under every product, improvements to detection reach the API, the investigation workspace and call protection together. ARCAS was trained on more than 10 million images and evaluated on more than 20 million images across over 5,000 generators in internal testing.
Teams that need to trust a face, voice or image they receive remotely.
Common uses include eKYC and digital onboarding, HR tech and remote hiring, account recovery, claims and refunds, and fraud and trust-and-safety investigation.
No. A low-risk result is not proof of authenticity.
It means the evaluated model found limited evidence of manipulation for that supported input under the conditions available. Unknown attacks, quality loss, missing context, and unsupported media can still affect the result.
A result can indicate technical signals that support a media-authenticity review.
It cannot independently identify the creator, infer intent, establish fraud, or determine whether an event, caption, or claim is true. Reviewers should combine it with provenance, source context, and other controls.
No. The product result is an input to a customer-controlled decision.
The customer sets thresholds, fallback behavior, manual-review rules, and the final action. Unsupported, rejected, or unavailable results need an explicit path rather than a silent approval.
The deepfake detection API for eKYC, onboarding and hiring.
Six signals in one request: Visual Safety Screening, Live Presence Verification, Face Manipulation Detection, AI Generative Model Analysis, AI Pixel-Level Manipulation and AI Content Compliance.
Each signal looks for a different kind of evidence. NeuroVerify combines them into one risk score and keeps the signal-level detail, so a reviewer can see which signal raised the result.
Yes. AI Generative Model Analysis looks for the traces generators leave, and AI Pixel-Level Manipulation looks for regions that were edited or regenerated.
This covers fully synthetic faces as well as real photos that were altered with AI tools, alongside face-swap and liveness checks in the same call.
Anywhere a person proves who they are with a selfie, a video or their voice.
Typical checkpoints are eKYC and customer onboarding, remote hiring and HR tech, account recovery and high-risk changes, at the volume those workflows run at.
Under 200 ms at p95 on eligible image requests, measured under defined test conditions.
That is fast enough to run inside a live onboarding or capture flow. The test methodology is available on request.
No. NeuroVerify analyses faces, video and audio, not identity documents.
Use it alongside your document-verification step. For investigating suspicious documents, NeuroDetect supports document evidence in an analyst-led case.
Choose the product that fits the media and review workflow.
Choose the product by media source and the decision workflow.
Use NeuroVerify for an application or API checkpoint, NeuroDetect for analyst-led investigation, and NeuroCall (Beta) for selected live-call workflows.
Face images, video and audio, through one API.
Video results can include separate visual and audio-track signals. Document investigation belongs to NeuroDetect.
NeuroDetect is a live, analyst-led workspace for investigating suspicious images, audio, video, and documents.
It keeps the source media, linked findings, model context, uncertainty, and analyst notes together for review. The analyst owns the conclusion.
NeuroCall is a Beta product that detects voice deepfakes in selected live calls.
Call-stream connections, deployment boundaries and response workflows are scoped with you for the intended environment.
Read scores, status fields, uncertainty, and supporting evidence.
A result means analysis completed; a rejection means the input was not scored as submitted; an error means analysis could not complete.
Keep these states separate in policy and reporting. A rejected or failed request is not a low-risk result and should follow a recapture, retry, fallback or review path.
Treat it as a model result for the submitted media, not as a statement of fact about the person or event.
Set workflow thresholds after testing representative genuine, manipulated, degraded and unsupported samples. Field meanings and score bands are shared with your account.
Yes. The available evidence changes with the input quality.
Compression, cropping, re-encoding, blur, noise, poor lighting, short duration, and missing metadata can weaken or alter useful signals. Test the conditions that occur in the real workflow.
A supported explainability view can direct a reviewer to a region, frame, interval, page, or signal.
It helps the reviewer inspect why a result was raised. It does not independently prove authorship, intent, source identity, or the truth of the depicted event.
Plan against the current public contract and customer-owned policy.
Start by signing in to Neural Defend for an API key, then call NeuroVerify from your backend at the point your workflow makes its decision.
Keep credentials on the server, preserve transaction references where required, and define what happens for every returned state: result, rejection and error.
Mobile, web and backend: Android (Kotlin), iOS (Swift), Flutter and web apps, plus Python, TypeScript/Node.js, Go and Java on the server.
Server SDKs run on Windows, macOS and Ubuntu or other Linux distributions. A REST API covers any other stack, and an MCP server lets AI agents call NeuroVerify directly.
Not by default. Automation should follow a tested, customer-owned policy.
Define thresholds, evidence requirements, recapture or fallback checks, manual-review capacity, and appeal handling. The right action depends on the cost of false positives and false negatives in the protected workflow.
Formats, file sizes and rate limits are set per endpoint and shared with your account when you sign in.
Image, video and audio inputs have their own quality rules and limits. We confirm the values that apply to your plan and workflow during implementation.
Confirm processing, access, retention, and operating boundaries.
Managed cloud API, private cloud or VPC, secure container, on-premises and air-gapped.
The managed API is the fastest way to start. Private, on-premises and air-gapped deployments are scoped during evaluation for your infrastructure, connectivity and update process.
Yes. On-premises and air-gapped deployments are supported and scoped during evaluation.
We confirm the runtime, infrastructure, capacity, update process and support access for your environment before rollout.
The answer must come from the selected product, deployment boundary, and agreed service terms.
Before use, document what media and metadata cross the boundary, where they are processed, what is retained, who can access logs or support evidence, and how deletion, residency, and incident handling work.
ISO/IEC 27001, ISO/IEC 27701, ISO/IEC 42001 and SOC 2 Type II.
Reports and certificates are available during security and procurement review, including scope and report period.
Test representative evidence before making an adoption decision.
There is no single accuracy number that describes every product, model version, media source, attack, threshold, and deployment.
Evaluate the relevant product with authorized samples that resemble the intended workflow. Report the test conditions, class balance, thresholds, false positives, false negatives, rejected inputs, errors, and uncertainty alongside any summary metric.
Define the decision first, then test representative inputs and operational failures.
Include genuine, manipulated, AI-generated, degraded, unsupported, and inconclusive samples. Compare the current control under the same conditions and test latency, throughput, timeouts, errors, reviewer evidence, and fallback behavior.
Measure the errors and operating conditions that affect the real decision.
Useful measures can include precision, recall, false-positive and false-negative behavior, completion and failure rates, latency, throughput, and reviewer workload. Record the threshold and test-set mix because both change how results should be interpreted.
Understand what Live, Beta, and Coming Soon mean today.
NeuroVerify is Live, NeuroDetect is Live, NeuroCall is Beta, and NeuroMeeting is Coming Soon.
These labels describe the current product stage. They do not by themselves confirm a particular integration, deployment model, geography, capacity, or production service term.
Live identifies a current product surface, Beta requires a scoped product discussion, and Coming Soon describes planned capability.
A Beta or planned interface should not be represented as generally available. A Live product can still require evaluation and availability confirmation for a specific environment or deployment boundary.
Not automatically.
Production fit depends on the supported input, integration contract, security review, capacity testing, failure handling, deployment boundary, and agreed operating terms for your workflow.
Source context
These answers follow the linked product, technology, deployment and evaluation pages. Product-specific limits can change, so confirm them with us during implementation.
