Liveness

The liveness algorithm enables LUNA PLATFORM to detect presentation attacks. A presentation attack is a situation when an imposter tries to use a video or photos of another person to circumvent the recognition system and gain access to the person’s private data.

There are the following general types of presentation attacks:

• Video Replay Attack. A video of another person is used

• Printed Photo Attack. One or several photos of another person are used

• Printed Mask Attack. An imposter cuts out a face from a photo and covers his face with it

• 3D Mask Attack. A silicone or plastic mask is used

Liveness check results

The liveness algorithm uses a single image for processing and returns the following data:

• Liveness probability [0..1].

The parameter shows the probability of a live person being present in the image, i.e. it is not a presentation attack. In general, the estimated probability must exceed the theoretical threshold of 50%. The value may be increased according to your business rules

• Image quality [0..1].

The parameter describes the integral value of image, facial, and environmental characteristics. In general, the estimated quality must exceed the theoretical threshold of 50%. The threshold may be increased according to the photo shooting conditions

See the “Liveness” section of the OpenAPI documentation for details.

[1]

APCER (Attack Presentation Classification Error Rate) — the rate of undetected attacks where algorithms identified the attack as a real person

[2] (1,2,3,4)

BPCER (Bona Fide Presentation Classification Error Rate) — the rate of incorrectly identified people where algorithms identified real people as fakes

Additional request parameters

You can specify the device OS type in the “OS” field of the “meta” object in the request:

• IOS

• ANDROID

• DESKTOP

• UNKNOWN

The parameter can decrease the overall error rate.

Requirements

There are certain requirements for image quality and face alignment that must be met to get correct results.

Face requirements

• A face should be fully open without any occlusions. The more face area is occluded, the lower the liveness estimation accuracy

• A face should be fully visible within a frame and should have padding around (the distance between the face and the image boundaries). The default minimum value of padding is 25 pixels. Cropping is not allowed

• Yaw and pitch angles are no more than 20 degrees in either direction

• The roll angle no more than 30 degrees in either direction

• The minimal distance between the eyes ~90 degrees (it is forbidden to set the value lower than 80 degrees)

• Single face in the image. It is recommended to avoid several faces being present in the image.

• No sunglasses

Capture requirements

• No blur (increases BPCER [2])

• No texture filtering (increases APCER [1])

• No spotlights on the face and close surroundings (increases BPCER [2])

• No colored light (increases BPCER [2])

• The face in the image must not be too light or too dark (increases BPCER [2])

• No fish-eye lenses

Image requirements

• Horizontal and vertical oriented images of 720p and 1080p

• Minimal image height: 480

• No or minimal image compression. The compression highly influences liveness algorithms