29、多通道人脸呈现攻击检测：方法、实验与结果

多通道人脸呈现攻击检测：方法、实验与结果

1. 训练框架与算法

在训练阶段，仅使用来自真实类别的嵌入来训练单类高斯混合模型（GMM）。测试时，样本首先通过网络得到嵌入 $x$，然后输入到单类 GMM 中得到对数似然分数：
[score = \log(p(x|\Theta))]

以下是训练框架的算法：

Algorithm 1: Algorithm for training the proposed framework
Data: (xi, yi), where xi is multi-channel input and yi ∈ {0, 1}; 0 – for attack and 1 – for bonafide
Result: WC – CNN weights, Θ_GMM – Parameters of GMM
1 Constants : λ – weighting factor, μ – learning rate
2 Initialize : CBF – center of bonafide class, WC – initial weights of CNN from pre-trained model
3 for mini-batch ← 1 to P do
4     Forward xi through the CNN
5     Compute the combined loss: L = (1 − λ)L_BCE + λL_OCCL
6     Back-propagate the loss and update the weights of DSUs and FC layers
7     U