Generalizable Representation Learning for Mixture Domain Face Anti-Spoofing

Authors: Zhihong Chen, Taiping Yao, Kekai Sheng, Shouhong Ding, Ying Tai, Jilin Li, Feiyue Huang, Xinyu Jin1132-1139

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that the proposed method outperforms conventional DG-based face anti-spoofing methods, including those utilizing domain labels.
Researcher Affiliation Collaboration 1College of Information Science & Electronic Engineering, Zhejiang University, 2 Youtu Lab, Tencent
Pseudocode Yes Algorithm 1 The optimization strategy of our D2AM
Open Source Code No The paper does not provide explicit links to open-source code or state that code will be released.
Open Datasets Yes Four public face anti-spoofing datasets are utilized to evaluate the effectiveness of our method: OULUNPU (Boulkenafet et al. 2017) (denoted as O), CASIAFASD (Zhang et al. 2012) (denoted as C), Idiap Replay Attack (Chingovska, Anjos, and Marcel 2012) (denoted as I), and MSU-MFSD (Wen, Han, and Jain 2015) (denoted as M).
Dataset Splits No The paper mentions 'meta-train' and 'meta-test' domains but does not explicitly describe a separate validation split or its specific parameters (percentages, counts, or methodology) for hyperparameter tuning in the main text.
Hardware Specification No The paper does not specify the hardware used for experiments, such as GPU models or CPU specifications.
Software Dependencies No Our method is implemented via Py Torch and trained with Adam optimizer. No version numbers for PyTorch or Adam are specified.
Experiment Setup Yes The learning rates α, β are set as 1e3, 1e-4, respectively, and the prior distribution for MMD is defined as the standard normal distribution. For other hyperparameters, we set λp as 0.1 and λm as 0.05. In our method, K determines the number of subdomains that the model needs to be divided. We found that converting the convolutional features extracted by pre-trained Res Net into domain features for clustering can clearly divide the sample into several clusters, so we can determine the value of K as 3.