3D Face Point Cloud Presentation Attack (3D-PCPA) Database

Description:

3D Face Point Cloude Presentation Attack (3D-PCPA) database acquired using an iPhone 12 Pro Smartphone. The frontal camera of the iPhone 12 Pro, in which the user interacts and captures the 3D point clouds on their own. The 3D scan was self-captured by the user at 15-20 cms between the face and smartphone. The 3D-PCPA database comprises bona fide and presentation attack face 3D point clouds to capture the point clouds acquired in multiple sessions.

Bona fide subset of 3D-PCPA

The bonafide face point clouds of the 3D-PCPA database were collected from 30 different data subjects in an indoor office environment. Each subject was asked to scan his/her own face using a smartphone in multiple sessions, and data collection was performed over 1-3 weeks’ time. A total of 1014 face point clouds were acquired, which corresponded to 30 to 33 3D point cloud samples per data subject. .

Presentation attack subsets for 3D-PCPA:

We considered two different 3D PAIs: (a) a 3D silicone mask and (b) a 3D wrap paper photo attack. The 3D silicone masks used in this work are custom-made high-quality face masks that have higher vulnerabilities to Face Recognition systems (FRS). Wrap paper photo attacks are generated by wrapping the print photo on the attacker’s face to simulate the pseudo depth.

3D Silicon Face Mask Artefact: We have employed four unique 3D silicone face mask that are used to capture the 3D point clouds. These 3D silicone masks were worn by the data subjects, and scans were performed on their own to capture the 3D point clouds. Data collection was carried out in different sessions for a duration of three weeks, where four masks were worn by 15 different data subjects. In total, the 3D silicon PAI had 840 point cloud scans corresponding to four unique masks.

3D Wrap photo Artefacts: The 3D wrap photo print attacks generation has three steps (1) We capture the high resolution photo of 15 different data subjects using DSLR camera (2) Digital photo is printed on a high quality papers using color laser printer (Model:Konica Minolta's bizhub C360i). (3) The printed photo was then wrapped around the face and self-captured using the frontal camera of the smartphone to obtain point clouds. The attacks were generated by wearing these 15 unique wraps by 20 different data subjects in multiple sessions varying from 1 to 3 weeks, resulting in 1626 3D wrap photo artefacts.

Following figure shows an example of 3D point cloudes samples from 3D-PCPA dataset

The statistics of 3D-PCPA database are as follows:

Researcher can avail 3D Face Point Cloud Presentation Attack (3D-PCPA) database by following the procedure mentioned:

Researchers are required to send the request to used 3D-PCPA database by sending the copy of licence agreement, completely filled to vetrekarnarayan@unigoa.ac.in with the subject line "License agreement for 3D Face Point Cloud Presentation Attack (3D-PCPA) database ".

Note: The license agreement has to be signed by the researcher or supervisor and the signature of legal authority on behalf of the institution, such as the Head of the institution or Registrar along with the institutional seal. The licence agreement should be on the researchers institutional letter head.

The request will not be considered if due procedures are not followed.

Every request to avail 3D-PCPA database will be placed before the Ethical Committee of the Institute for the approval and the confirmation of request will be sent via email. Further, all the instructions related to access to database will be detailed in the email.

This database is available only for research and educational purpose and not for any commercial use. Further, the use of database is only for the period of one year from the data of signing the agreement from the researcher. Beyond the allocated one year for the usage of data, the applicant have to apply freshly again following the due procedure.

Database is available only for research and educational purpose only. All the rights of the 3D Face Point Cloud Presentation Attack (3D-PCPA) database are reserved and commercial use/distribution of this database is strictly prohibited. All the technical reports and papers that report experimental results from this database should provide due acknowledgement and reference. If you use the database in any publications or reports, you must refer to the following paper:

Raghavendra Ramachandra, Narayan Vetrekar, Sushma Venkatesh, Jag Mohan Singh, Savita Nageshker, R. S. Gad, VoxAtnNet: A 3D Point Clouds Convolutional Neural Network for Generalizable Face Presentation Attack Detection, 18th IEEE International Conference on Automatic Face and Gesture Recognition (FG), Turkey, pp.1-9, 2024.

3D Face Point Cloud Presentation Attack (3D-PCPA) V2 Database

Large Scale Database will be available soon

Description:

3D Face Point Cloud Presentation Attack (3D-PCPA) V2 Database was collected using two different smartphones. The 3D-PCPA V2 database is an expansion of the 3D-PCPA database and was obtained using iPhone11 Pro and iPhone 12 ProMax smartphones. The 3D scan was self-captured by the user at 15-20 cms between the face and smartphone. The 3D-PCPA database comprises bona fide and presentation attack face 3D point clouds to capture the point clouds acquired in multiple sessions. The 3D-PCPA V2 database was divided into two subsets: bona fide and PAI Face 3D point cloud data, which were collected during multiple sessions over an eight-month period. The database obtained 4557 and 4624 3D face point clouds using the iPhone 11 Pro (we refer to this as P1) and the iPhone 12 ProMax (we refer to this as P2), respectively.

Bona fide subset of 3D-PCPA V2 Database

The Bona fide subset of 3D-PCPA V2 consists of a 3D face point cloud corresponding to 3693 and 3872 3D scans collected using P1 and P2, respectively. In total, 100 data points were captured, of which 50 corresponded to males and 50 corresponded to females. For each subject, a minimum of 20-30 face point clouds were acquired in multiple sessions over an eight months duration. The 3D face point clouds were self-captured by the users independently using both the smartphones resulting in a total of 7565 bona fide face point clouds.

Presentation attack subsets for 3D-PCPA:

To present the variability of different types of PAI in PAD techniques, we employed six different types of PAIs to acquire 3D point clouds. Presentation attacks are presented to smartphones by wearing PAI masks in conjunction with disguise accessories such as wigs and hats to conceal the discontinuities of the PAI and present a more realistic appearance. The presentation attack samples were collected in multiple sessions over a period of eight months.

Hard Plastic Mask Artefacts: This PAI is constructed using long-lasting plastics that are often fashioned to resemble the features of a human face or fictional characters, thereby granting the attacker a certain level of anonymity and disguise. We used three different colored hard plastic masks (Red, Green, Blue), along with two different types of wigs and hats, to generate three different variants of presentation attacks. Three variants were generated using different combinations of hard plastic, wig, and hat, which were worn by ten different data subjects to self-capture presentation attack 3D face point clouds in multiple sessions with a duration of eight months. We label this PAI as Attack #1 or A1.

Latex Mask Artefacts: A custom-made 3D face mask constructed from latex material was utilized in this study due to its ability to closely replicate human facial features, providing a high level of disguise. Three distinct 3D latex face mask artifacts, comprising two males and one female, were employed to generate 3D point clouds. To enhance realism, a wig was also incorporated, and the masks were worn by 15 different users across multiple sessions spanning a period of five to seven months to capture the necessary data. This PAI is referred to as Attack #2 or A2.

Paper Mask Artefacts: A human-shaped face mask constructed from durable paper pulp cardboard, which is both thick and stiff, was utilized in this work. We used two white paper masks (one male and one female) that were blank and not painted to generate artefacts. Different combinations of two paper masks and two wigs were worn by eight different data subjects to create four different variants of attacks. This PAI is referred to as Attack #3 or A3.

Silicon Face Mask Artefact: The silicone mask artefacts in the 3D-PCPA V2 dataset are the same as those of the 3D-PCPA V1 dataset introduced in our earlier work [34]. A custom-made high-quality 3D silicon face mask with greater vulnerability to face recognition systems (FRS) was employed in this work. We considered four silicon face masks, two males and two females, to capture 3D point cloud data. These masks were worn by the data subject along with the wig to present a realistic attack on the FRS. A total of 15 data subjects participated in wearing the silicon face mask to self-capture the presentation attack samples. This PAI is referred to as Attack #4 or A4

Soft Plastic Mask Artefacts: To capture the presentation attack samples from this artefact, we employed a lightweight flexible face craft molded like a face mask to closely resemble a human face. We used a single soft plastic mask along with two different wigs, which were worn by eight different data subjects to self-capture 3D point clouds using the front cameras of both smartphones. This PAI is referred to as Attack #5 or A5.

Wrap Photo Artefacts: The wrap face mask attack in the 3D-PCPA V2 dataset is the same as that in the 3D-PCPA V1 dataset introduced in our earlier work [34]. Color-printed photographs of the Vitim data subjects were used to create a photo wrap attack with a pseudo-depth effect. This process involved three steps: (a) High-resolution color photographs of 15 different subjects were acquired using a 24-megapixel DSLR camera (Model: Nikon D320). Ten of the subjects were male and five were female. (b) Digital photos were printed on high-quality paper using a laser printer (Model: Konica Minolta’s Bizhub C360i). (c) The printed photograph is then wrapped around the face along with a wig to capture the pseudo-depth face point cloud using the front true depth camera of a smartphone. This PAI is referred to as Attack #6 or A6.

Following figure shows an example of 3D point cloud samples from 3D-PCPA V2 database

The statistics of 3D-PCPA V2 database summarizing the data partition for training and testing set is as follows:

Proposed 3D Face PAD:

Block diagram of the proposed PCGattnNet for face PAD is as follows.

The innovative aspect of PCGattnNet lies in its utilization of the graph attention mechanism to represent point clouds dynamically. This was further enhanced by the dual-stream architecture, which captures discriminant information suitable for a generalizable PAD. The implementation of PCGattnNet is available to download.

Researcher can avail 3D Face Point Cloud Presentation Attack (3D-PCPA) V2 database by following the procedure mentioned:

Researchers are required to send the request to used 3D-PCPA V2 database by sending the copy of licence agreement, completely filled to vetrekarnarayan@unigoa.ac.in with the subject line "License agreement for 3D Face Point Cloud Presentation Attack (3D-PCPA) V2 database ".

The request will not be considered if due procedures are not followed.

Every request to avail 3D-PCPA V2 database will be placed before the Ethical Committee of the Institute for the approval and the confirmation of request will be sent via email. Further, all the instructions related to access to database will be detailed in the email.

Database is available only for research and educational purpose only. All the rights of the 3D Face Point Cloud Presentation Attack (3D-PCPA) V2 database are reserved and commercial use/distribution of this database is strictly prohibited. All the technical reports and papers that report experimental results from this database should provide due acknowledgement and reference. If you use the database in any publications or reports, you must refer to the following paper:

Raghavendra Ramachandra, Narayan Vetrekar, Sushrut Patwardhan, Sushma Venkatesh, Gauresh Naik and R. S. Gad, "PCGattnNet: A 3D Point Cloud Dynamic Graph Attention for Generalizable Face Presentation Attack Detection," in IEEE Transactions on Biometrics, Behavior, and Identity Science, doi: 10.1109/TBIOM.2025.3534641, 2025

Page updated

Google Sites

Report abuse