Custom Silicone Mask Attack Dataset (CSMAD)

The Custom Silicone Mask Attack Dataset (CSMAD) contains presentation attacks made of six custom-made silicone masks. Each mask cost about USD 4000. The dataset is designed for face presentation attack detection experiments. The Custom Silicone Mask Attack Dataset (CSMAD) has been collected at the Idiap Research Institute. It is intended for face presentation attack detection experiments, where the presentation attacks have been mounted using a custom-made silicone mask of the person (or identity) being attacked. The dataset contains videos of face-presentations, as a set of files specifying the experimental protocol corresponding the experiments presented in the corresponding publication. Reference If you publish results using this dataset, please cite the following publication. Sushil Bhattacharjee, Amir Mohammadi and Sebastien Marcel: "Spoofing Deep Face Recognition With Custom Silicone Masks." in Proceedings of International Conference on Biometrics: Theory, Applications, and Systems (BTAS), 2018. 10.1109/BTAS.2018.8698550 http://publications.idiap.ch/index.php/publications/show/3887 Data Collection Face-biometric data has been collected from 14 subjects to create this dataset. Subjects participating in this data-collection have played three roles: targets, attackers, and bona-fide clients. The subjects represented in the dataset are referred to here with letter-codes: A .. N. The subjects A..F have also been targets. That is, face-data for these six subjects has been used to construct their corresponding flexible masks (made of silicone). These masks have been made by Nimba Creations Ltd., a special effects company. Bona fide presentations have been recorded for all subjects A..N. Attack presentations (presentations where the subject wears one of 6 masks) have been recorded for all six targets, made by different subjects. That is, each target has been attacked several times, each time by a different attacker wearing the mask in question. This is one way of increasing the variability in the dataset. Another way we have augmented the variability of the dataset is by capturing presentations under different illumination conditions. Presentations have been captured in four different lighting conditions: flourescent ceiling light only halogen lamp illuminating from the left of the subject only halogen lamp illuminating from the right only both halogen lamps illuminating from both sides simultaneously All presentations have been captured with a green uniform background. See the paper mentioned above for more details of the data-collection process. Dataset Structure The dataset is organized in three subdirectories: ‘attack’, ‘bonafide’, ‘protocols’. The two directories: ‘attack’ and ‘bonafide’ contain presentation-videos and still images for attacks and bona fide presentations, respectively. The folder ‘protocols’ contains text files specifying the experimental protocol for vulnerability analysis of face-recognition (FR) systems. The number of data-files per category are as follows: ‘bonafide’: 87 videos, and 17 still images (in .JPG format). The still images are frontal face images captured using a Nikon Coolpix digital camera. ‘attack’: 159, organized in two sub-folders – ‘WEAR’ (108 videos), and ‘STAND’ (51 videos) The folder ‘attack/WEAR’ contains videos where the attack has been made by a person (attacker) wearing the mask of the target being attacked. The ‘attack/STAND’ folder contains videos where the attack has been made using a the target’s mask mounted on an appropriate stand. Video File Format The video files for the face-presentations are in ‘hdf5’ format (with file-extensions ‘.h5’. The folder structure of the hdf5 file is shown in Figure 1. Each file contains data collected using two cameras: RealSense SR300 (from Intel): collects images/videos in visible-light (RGB color) , near infrared (NIR) @ 860nm wavelength, and depth maps Compact Pro (from Seek Thermal): collects thermal (long-wave infrared (LWIR)) images. As shown in Figure 1, frames from the different channels (color, infrared, depth, thermal) from he two cameras are stored in separate directory-hierarchies in the hdf5 file. Each file respresents a video of approximately 10 seconds, or roughly, 300 frames. In the hdf5 file, the directory for SR300 also contains a subdirectory named ‘aligned_color_to_depth’. This folder contains post-processed data, where the frames of depth channel have been aligned with those of the color channel based on the time-stamps of the frames. Experimental Protocol The ‘protocols’ folder contains text files that specify the protocols for vulnerability analysis experiments reported in the paper mentioned above. Please see the README file in the protocols folder for details.

自定义硅胶面具攻击数据集（Custom Silicone Mask Attack Dataset, CSMAD）包含6个由定制硅胶面具制作的人脸呈现攻击样本，单张面具的制作成本约为4000美元。本数据集专为**人脸呈现攻击检测**实验设计，由Idiap研究院采集完成，旨在开展基于定制硅胶面具的人脸呈现攻击检测研究，攻击手段为使用目标对象（或目标身份）的专属定制硅胶面具实施攻击。数据集包含人脸呈现视频，以及对应已发表论文中实验方案的配套文件。

### 参考文献
若使用本数据集发表研究成果，请引用以下论文：
Sushil Bhattacharjee、Amir Mohammadi与Sebastien Marcel：《Spoofing Deep Face Recognition With Custom Silicone Masks》，收录于2018年国际生物识别理论、应用与系统会议（International Conference on Biometrics: Theory, Applications, and Systems, BTAS），DOI: 10.1109/BTAS.2018.8698550，原文链接：http://publications.idiap.ch/index.php/publications/show/3887

### 数据采集
本数据集从14名受试者处采集人脸生物特征数据。参与本次采集的受试者承担三类角色：目标对象、攻击者与真实合法用户。数据集中的受试者以字母代码A至N指代。其中A至F共6名受试者同时作为目标对象，其人脸数据被用于制作对应的柔性硅胶面具，该批面具由特效公司Nimba Creations Ltd.定制。

所有受试者A至N均录制了真实合法的人脸呈现样本。针对6名目标对象，均录制了由不同受试者佩戴对应面具实施的攻击呈现样本——即每名目标对象均接受多名不同攻击者的面具攻击，以此提升数据集的样本变异性。此外，为进一步丰富数据集多样性，采集环节设置了四种不同的光照条件：仅荧光顶灯照明、仅卤素灯从受试者左侧照明、仅卤素灯从受试者右侧照明、双侧卤素灯同时照明。所有采集场景均采用绿色纯色背景。更多采集细节请参阅前文提及的研究论文。

### 数据集结构
本数据集分为三个子目录："attack"、"bonafide"与"protocols"。其中"attack"与"bonafide"目录分别存放攻击呈现与真实合法呈现的视频及静态图像文件；"protocols"目录包含用于**人脸识别（Face Recognition, FR）**系统脆弱性分析的实验方案文本文件。各类别数据文件数量如下：
- "bonafide"：87段视频与17张静态图像（格式为JPEG（Joint Photographic Experts Group）），静态图像为使用尼康Coolpix数码相机拍摄的正面人脸照片。
- "attack"：共159段视频，分为"WEAR"（108段）与"STAND"（51段）两个子文件夹。"attack/WEAR"目录存放攻击者佩戴目标对象面具实施攻击的视频；"attack/STAND"目录存放将目标对象面具固定于专用支架上实施攻击的视频。

### 视频文件格式
人脸呈现的视频文件采用HDF5（Hierarchical Data Format 5）格式，文件扩展名为".h5"，其文件夹结构如图1所示。每个视频文件由两台相机的采集数据构成：
1. 英特尔RealSense SR300相机：采集可见光（RGB彩色）图像、860nm波长近红外（Near Infrared, NIR）图像与深度图
2. Seek Thermal Compact Pro相机：采集热成像（长波红外（Long-Wave Infrared, LWIR））图像

如前文图1所示，两台相机的不同通道（彩色、红外、深度、热成像）的帧数据分别存储于HDF5文件的独立目录层级中。每个视频文件时长约10秒，包含约300帧图像。在HDF5文件中，SR300相机的目录下还包含名为"aligned_color_to_depth"的子目录，该文件夹存放经过后处理的数据，即根据帧时间戳将深度通道图像帧与彩色通道图像帧进行对齐。

### 实验方案
"protocols"目录包含的文本文件定义了前文提及论文中报道的脆弱性分析实验方案。更多细节请参阅"protocols"目录下的README文件。

搜集汇总

背景与挑战

背景概述

Custom Silicone Mask Attack Dataset (CSMAD) 是一个专为面部呈现攻击检测设计的数据集，使用6个定制硅胶面具（每个成本约4000美元）模拟攻击，增强现实世界攻击场景的逼真度。数据集包含14名受试者的多模态数据（如可见光、近红外、深度和热成像视频），并在四种不同光照条件下收集，以增加变异性，适用于面部识别系统的脆弱性分析实验。数据量包括87个真实呈现视频和159个攻击呈现视频，结构清晰，支持协议化实验。