81、一致性正则化助力缓解鲁棒过拟合

一致性正则化助力缓解鲁棒过拟合

1. 算法介绍

1.1 PGD - AT+MT 算法

以下是 PGD - AT+MT 算法的详细步骤：

Algorithm 1. PGD - AT+MT
Require: training set D; batch size m; learning rate lr; number of training epochs T ; PGD step size
α; number of PGD steps K; max perturbation budget ϵ; EMA smoothing parameter η; EMA
start epoch Es
Ensure: Robust model weight θt
1: Randomly initialize θs
2: for t = 1 to T do
3:
    for i = 1 to m do
4:
        Sample xi from D
5:
        x′
        i ←xi + ϵδ, where δ ∼Uniform(−1, 1)
6:
        for k = 1 to J do
7:
            if t < Es then
8:
                x′
                i ←Πϵ

                x′
                i + α sign

                ∇x′
                iLce

                f

                x′