Do Deepfake Detectors Have Demographic Bias? Subgroup Performance Evaluation

Do deepfake detectors have demographic bias?

They can, and the research shows it is common in models trained on narrow data. Loc Trinh and Yan Liu examined three popular deepfake detectors in a paper accepted to IJCAI 2021 and found error rates that differed by up to 10.7 percent between racial subgroups. They traced part of the problem to training data: FaceForensics++, one of the most used benchmarks, was overwhelmingly made up of Caucasian subjects, and detectors trained on blended-image data showed systematic errors against female Asian faces.

A larger study by Ying Xu and colleagues, first posted in 2022 and revised in 2024, added 47 demographic and non-demographic attribute labels to five popular deepfake datasets, over 65 million labels in total. It found limited diversity in the datasets and found three detector backbones strongly affected by the attributes studied, from age, gender and ethnicity to hair, skin and accessories.

Face recognition has a longer record. NIST's December 2019 report NISTIR 8280 found false positive rates that varied by a factor of up to 100 between demographic groups for many face recognition algorithms. Deepfake detection is a different task, but it shares the same exposure to unbalanced training data.

Why subgroup gaps matter in a deepfake detection workflow

An average error rate can hide a group that bears most of the errors. If a detector flags 1 percent of real selfies overall but 4 percent from one group, those customers face four times the friction, delay and suspicion, and the business loses more of them. In regulated services, unequal treatment can also become a compliance and fairness issue that sits outside any cost model.

False negatives can be unequal too. If fakes built on certain faces pass more often, attackers will learn to use those faces. Both directions need measuring.

Gaps also move over time. A model update, a new capture SDK or a shift in your customer base can open a gap that the launch evaluation did not show. Treat subgroup results as something you monitor, with a named owner and a review date, rather than a one-off procurement check.

Which subgroups to measure, and how to get the labels

Pick groups from the people and conditions your service actually reaches, then measure the ones where an error would cause harm. Demographic labels need a lawful basis, consent and careful handling under data protection law, so many teams start with attributes they can record without collecting sensitive data.

ISO/IEC TR 24027:2021 describes sources of bias in AI systems and ways to assess it, and is a useful frame for agreeing which attributes to study.

How to run a subgroup performance evaluation

Sample enough items per group to detect the gap you care about. With 300 authentic selfies in a group, a 1 percent false positive rate means about three flags, which is too few to tell 1 percent from 3 percent with confidence. Groups often need over a thousand authentic items each before small gaps become measurable. Over-sample small groups rather than relying on their natural share of traffic.

Use one fixed threshold for all groups, then report false positive rate, false negative rate and inconclusive rate per group with confidence intervals. Report the largest ratio between groups as well as the differences, because a jump from 0.5 to 1.5 percent is a threefold gap even though it looks small.

Look at intersections where samples allow, such as older users on budget devices in low light. Gaps often appear in combinations that single-attribute tables miss.

What to do when you find a gap

First check whether the gap comes from the detector or from capture. Poor lighting and low-end cameras affect some groups more than others, and better capture guidance can close a gap without touching the model. If the gap persists, share the cases with the vendor and ask for retraining or recalibration on more diverse data, then retest on the same held-out groups.

In the meantime, route flags from affected groups to human review rather than automatic rejection, and monitor flag and override rates by group in production. Avoid per-group thresholds unless legal review approves them, since setting different rules by protected characteristic raises its own problems.

Subgroup evidence from Neural Defend

The ARCAS 1B paper, which documents the model behind our products, reports subgroup breakdowns alongside pooled results and keeps the two separate, so readers can see where performance varies rather than relying on one headline figure. Its authors also state that the results describe the records tested and do not establish general reliability for every population.

For your evaluation, NeuroVerify returns a unified risk score with signal-level evidence for every file, so you can compute error rates by group, device and condition on your own sealed set. We expect buyers to run that analysis before production and to repeat it after model updates.

Because NeuroVerify returns six signals rather than a single verdict, a subgroup gap can be traced to the signal that causes it, for example whether extra flags on low-light selfies come from face manipulation detection or from pixel-level analysis. That makes the conversation about fixes specific, and it gives your fairness review evidence it can document.