Liveness, Face Matching, Black List Checks

This article describes the main types of analyses that Oz software is able to perform.

• Liveness checks whether a person in a media is a real human.

• Face Matching examines two or more media to identify similarities between the faces depicted in them.

• Black list looks for resemblances between an individual featured in a media and individuals in a pre-existing photo database.

These analyses are accessible in the Oz API for both SaaS and On-Premise models. Liveness and Face Matching are also offered in the On-Device model. Please visit this page to learn more about the usage models.

Liveness

The Liveness check is important to protect facial recognition from the two types of attacks.

A presentation attack, also known as a spoofing attack, refers to the attempt of an individual to deceive a facial recognition system by presenting into a camera video, photo, or any other type of media that mimics the appearance of a genuine user. These attacks can include the use of realistic masks, or involve digital manipulation with images and videos, such as deep fakes.

An injection attack is an attempt to deceive a facial recognition system by replacing physical camera input with a prerecorded image or video, or by manipulating physical camera output before it becomes input to a facial recognition. Virtual camera software is the most common tool for injection attacks.

Oz Liveness is able to detect both types of attacks. Any component can detect presentation attacks, and for injection attack detection, use Oz Liveness Web SDK. To learn about how to use Oz components to prevent attacks, check our integration quick start guides:

• How to integrate server-based liveness into your web application,

• How to integrate server-based liveness into your mobile application,

• How to check your media for liveness without Oz front end.

Once the Liveness check is finished, you can check both qualitative and quantitative analysis results.

Results overview

Qualitative results

• SUCCESS – everything went fine, the analysis has completed successfully;

• DECLINED – the check failed (an attack detected).

If the analysis hasn't been finished yet, the result can be PROCESSING (the analysis is in progress) / FAILED (the analysis failed due to some error and couldn't get finished).

If you have analyzed multiple media, the aggregated status will be SUCCESS only if each analysis on each media has finished with the SUCCESS result.

Quantitative results

• 100% (1) – an attack is detected, the person in the video is not a real living person,

• 0% (0) – a person in the video is a real living person.

Asking users to perform a gesture, such as smiling or turning their head, is a popular requirement when recording a Liveness video. With Oz Liveness Mobile and Web SDK, you can also request gestures from users. However, our Liveness check relies on other factors, analyzed by neural networks, and does not depend on gestures. For more details, please check Passive and Active Liveness.

Liveness check also can return the best shot from a video: a best-quality frame where the face is seen the most properly.

Face Matching

The Biometry algorithm allows comparing several media and check if the people on them are the same person or not. As sources, you can use images, videos, and scans of documents (with photo). To perform the analysis, the algorithm requires at least two media.

Results overview

Qualitative results

• SUCCESS – everything went fine, the analysis has completed successfully;

• DECLINED – the check failed (faces don't match).

If the analysis hasn't been finished yet, the result can be PROCESSING (the analysis is in progress) / FAILED (the analysis failed due to some error and couldn't get finished).

Quantitative results

After comparison, the algorithm provides numbers that represent the similarity level. The numbers vary from 100 to 0% (1 to 0), where:

• 100% (1) – faces are similar, media represent the same person,

• 0% (0) – faces are not similar and belong to different people

There are two scores to consider: the minimum and maximum. If you have analyzed two media, these scores will be equal. For three or more media, the similarity score is calculated for each pair. Once calculated, these scores get aggregated and analysis returns the minimum and maximum similarity scores for the media compared. Typically, the minimum score is enough.

Wonder how to integrate face matching into your processes? Check our integration quick start guides.

Black List

In Oz API, you can configure one or more black lists, or face collections. These collections are databases of people depicted in photos. When the Black list analysis is being conducted, Oz software compares the face in a photo or video taken with faces of this pre-made database and shows whether a face exists in a collection.

Results overview

Qualitative results

• SUCCESS – everything went fine, the analysis has completed successfully;

• DECLINED – the check failed (faces match).

If the analysis hasn't been finished yet, the result can be PROCESSING (the analysis is in progress) / FAILED (the analysis failed due to some error and couldn't get finished).

Quantitative results

After comparison, the algorithm provides a score that represents the similarity level. The number varies from 100 to 0% (1 to 0), where:

• 100% (1) – the person in an image or video matches with someone in the blacklist database,

• 0% (0) – the person is not found in the blacklist.

For additional information, please refer to this article.