Spoof Detection

Active Spoof

Active spoof works by observing the geometry and perspective of a face as it moves closer to a camera. Active spoof requires two input images, one far and one close, and is an involved process. Therefore, please be sure to consult the sample code for a demonstration of proper usage.

SDK.check_spoof_image_face_size(self: tfsdk.SDK, tf_image: tfsdk.TFImage, face_box_and_landmarks: tfsdk.FaceBoxAndLandmarks, active_spoof_state: tfsdk.ACTIVESPOOFSTAGE) → tfsdk.ERRORCODE

Ensures that the face size meets the requirements for active spoof. Must check function return value! Active spoof works by analyzing the way a persons face changes as they move closer to a camera. The active spoof solution therefore expects the face a certain distance from the camera. In the far image, the face should be about 18 inches from the camera, while in the near image, the face should be 7-8 inches from the camera. This function must be called before calling tfsdk.SDK.detect_active_spoof().

Parameters

• tf_image - the input tfsdk.TFImage, returned by tfsdk.SDK.preprocess_image().

• face_box_and_landmarks - The face on which to run active spoof detection.

• active_spoof_state - The stage of the image, either near stage or far stage.

Returns

The ERRORCODE. If ERRORCODE.NO_ERROR is returned, then the image is eligible for active spoof detection. If ERRORCODE.FACE_TOO_CLOSE or ERRORCODE.FACE_TOO_FAR is returned, the image is not eligible for active spoof detection.

SDK.detect_active_spoof(self: tfsdk.SDK, near_face_landmarks: Annotated[List[tfsdk.Point], FixedSize(106)], far_face_landmarks: Annotated[List[tfsdk.Point], FixedSize(106)]) → Tuple[tfsdk.ERRORCODE, float, tfsdk.SPOOFLABEL]

Detect if there is a presentation attack attempt. Must call tfsdk.SDK.check_spoof_image_face_size() on both input faces before calling this function.

Parameters

• near_face_landmarks - The face landmarks of the near face, obtained by calling tfsdk.SDK.get_face_landmarks().

• far_face_landmarks - The face landmarks of the far face, obtained by calling tfsdk.SDK.get_face_landmarks().

Returns

The ERRORCODE, the spoof score, and the tfsdk.SPOOFLABEL, in that order.

class tfsdk.ACTIVESPOOFSTAGE

Members:

NEAR : Near image.

FAR : Far image.

class tfsdk.SPOOFLABEL

Members:

FAKE : The image may be an attempted spoof.

REAL : The face image is real.

Passive Spoof

Passive spoof is meant to be used for selfie-style images from smartphones or from laptop webcams, and should not be used with CCTV cameras / cameras at a distance. Unlike active spoof which requires two input images (one close and one far), passive spoof requires only a single input image. This being said, there are still strict requirements for said input image:

• Face must be at least 100 pixels in height. Face height must also be between 0.323 and 0.55 of the total image height.

• The face must be centered vertically and horizontally within the frame for portrait images. Landscape images are more permissive for horizontal centering as a portrait region is cropped around the face.

• The face must have neutral yaw, pitch, and roll, and must be facing the camera directly.

• The eyes must be open.

• User must not be wearing a mask.

• The image must not be overly bright or dark.

If any of the above criteria are not met, the method will return an error code. It is therefore imperative to check the return value of the function. Please consult the sample code for proper usage of this method. The sample code also demonstrates how to draw an oval of the correct size, prompting the user to place their face in the oval for easy capture of a conforming image. We advise running this method on the front end so that immediate feedback can be provided to the user until they submit a conforming image. We also advise running multiple images through the spoof detector (4-6) and then taking the average result.

SDK.detect_spoof(*args, **kwargs)

Overloaded function.

1. detect_spoof(self: tfsdk.SDK, tf_image: tfsdk.TFImage, face_box_and_landmarks: tfsdk.FaceBoxAndLandmarks, threshold: float = 0.75) -> Tuple[tfsdk.ERRORCODE, tfsdk.SPOOFLABEL, float]

   Detect if there is a spoof attempt. Returns the spoof score. Passive spoof detection only works with “selfie-style”, meaning the face must be up close and take up most of the frame.

   param tf_image

   the input tfsdk.TFImage, returned by tfsdk.SDK.preprocess_image().

   param face_box_and_landmarks

   tfsdk.FaceBoxAndLandmarks returned by tfsdk.SDK.detect_faces() or tfsdk.SDK.detect_largest_face().

   param threshold

   the spoof score threshold above which it is considered a real image (default = 0.75). Refer to the ROC curves when selecting a threshold.

   return

   The ERRORCODE, the tfsdk.SPOOFLABEL, and the spoof score, in that order.

2. detect_spoof(self: tfsdk.SDK, tf_image: tfsdk.TFImage, face_box_and_landmarks: tfsdk.FaceBoxAndLandmarks, yaw: float, roll: float, blink_state: tfsdk.BlinkState, mask_label: tfsdk.MASKLABEL, threshold: float = 0.75) -> Tuple[tfsdk.ERRORCODE, tfsdk.SPOOFLABEL, float]

   Detect if there is a spoof attempt. Returns the spoof score. Passive spoof detection only works with “selfie-style”, meaning the face must be up close and take up most of the frame.

   param tf_image

   the input tfsdk.TFImage, returned by tfsdk.SDK.preprocess_image().

   param face_box_and_landmarks

   tfsdk.FaceBoxAndLandmarks returned by tfsdk.SDK.detect_faces() or tfsdk.SDK.detect_largest_face().

   param yaw

   the rotation angle around the image’s vertical axis, in radians, as returned by tfsdk.SDK.estimate_head_orientation().

   param roll

   the rotation angle around the image’s longitudinal axis, in radians, as returned by tfsdk.SDK.estimate_head_orientation().

   param blink_state

   the output of the blink prediction, as returned by tfsdk.SDK.blink_state().

   param mask_label

   the predicted MaskLabel for face image, as returned by tfsdk.SDK.detect_mask().

   param threshold

   the spoof score threshold above which it is considered a real image (default = 0.75). Refer to the ROC curves when selecting a threshold.

   return

   The ERRORCODE, the tfsdk.SPOOFLABEL, and the spoof score, in that order.

Below are two examples of conforming images, one landscape, and one portrait:

Below are four examples of non-conforming images:

Face too far:

Face too close:

Face yaw angle too extreme:

Face roll angle too extreme:

Image too bright: