Deepfake Detection API Latency and Throughput: How to Compare Vendors

How fast is deepfake detection and what latency should I expect?

For a still image, a fast commercial API returns a result in a few hundred milliseconds, and the slowest take many seconds. The only published side-by-side figures we know of come from VendorBench-100, a public benchmark of 36 deepfake image detection systems. Its v2 paper, released in July 2026, reported mean latency for the commercial APIs it tested on its own setup. The authors note these figures are not comparable across tracks, and they measured still images only.

Video and audio take longer, and the time depends on duration, resolution and how many frames or segments the service analyses. A short selfie video can be screened in seconds; an hour of call audio is a batch job. Ask each vendor how latency scales with media length, and test with files at the lengths your users actually submit.

What to measure: the latency your users actually feel

Vendors often quote model inference time. Your users experience something longer: upload from the device, network transit, queueing, any preprocessing, inference, the response, and your own decision logic. Define the boundary before testing, and measure from your server, in your region, through the same SDK or HTTP client you will use in production.

Report the distribution, not a single number. The median shows typical speed; p95 and p99 show the slow tail that drives timeouts and abandoned sessions. Track time to first byte separately for large uploads, and record cold-start behaviour after idle periods, since a serverless or autoscaled service can be slow on the first request after a quiet spell.

Then set a latency budget for the journey step. Add up what the step already spends on capture, upload and your own checks, decide how long a user will reasonably wait, and give detection what is left, minus room for one retry. If a vendor's p95 does not fit that budget at peak, the choice is between moving detection out of the live path or choosing a faster service, and it is better to know that before the contract than after launch.

How to load-test throughput and capacity

Model real demand first: average and peak requests per second, burst length, media mix and size distribution. Then run three tests. A steady test at expected peak for at least an hour shows stable latency and error rates. A step test raises load until p95 latency or errors cross your limit, which marks usable capacity. A spike test throws a sudden burst at the API and measures recovery.

Capacity planning follows a simple relationship, often called Little's Law: requests in flight equal throughput multiplied by latency. At 50 requests per second, an API answering in 0.2 s has about 10 requests open at once. An API answering in 2 s has about 100, which means more connections, larger rate limits and more memory on your side for the same traffic.

Count every error and timeout. Check how the API signals rate limiting, whether retries are safe, and what happens to queued requests during an outage.

• Steady load at expected peak: p50, p95, p99 latency and error rate.
• Step load to failure: the request rate where limits are crossed.
• Spike: recovery time and whether any requests are lost or duplicated.
• Mixed media: images, short videos and audio in production proportions.

Choosing synchronous, asynchronous or batch processing

Use synchronous calls when a person is waiting, such as a selfie check inside onboarding, and only when p95 latency fits inside the step's time budget with room for retries. Use asynchronous calls with webhooks for video, longer audio and anything where a delay of seconds is acceptable. Use batch jobs for back-office review, re-screening of past submissions and investigations.

Whatever the mode, decide what happens when the result is late. A timeout should hold the case or route it to review. It should never let the case pass by default.

Turning load test results into operating limits

Write the results into numbers your team can monitor: the maximum request rate per environment, the p95 latency target, the timeout and retry policy, and the alert thresholds that page someone. Agree service levels with the vendor that match those figures, and confirm that the support model covers your peak hours.

Retest after major vendor releases and before seasonal peaks. A model update that improves accuracy can also change speed.

NeuroVerify latency in practice

NeuroVerify completes eligible image requests in under 200 ms at p95 under defined test conditions, with the methodology available on request, and that single call returns all six signals rather than one. Speed of that order lets detection sit inside a live onboarding or account recovery step without a visible wait. VendorBench-100 v2 separately measured Neural Defend's mean still-image latency at 175 ms on its own setup.

Deployment options include managed cloud, private cloud or VPC, secure container, on-premises and air-gapped, so inference can run close to your traffic and inside your network where data rules require it. SDKs for Android, iOS, Flutter, Python, TypeScript and Node.js, Go and Java keep client overhead low. Load-test us at your own peak during evaluation; we will share the conditions behind our figures so you can reproduce them.