3.2 DP-SGD算法步骤 计算单样本的梯度 gi=∇ωL(xi,yi)g_i=\nabla_{\omega}\mathcal{L}(x_i, y_i)gi=∇ωL(xi,yi) 单样本梯度剪裁 gˉi=gi/max(1,∥gi∥2C)\bar{g}_i=g_i/\max\Big(1,\frac{\|g_i\|_2}{C}\Big)gˉi=gi/max(1,C∥gi...
For this reason, it would be much preferable if we could instead insert the DP mechanismduring model training, so that the resulting model could be safe for release. This brings us to the DP-SGD algorithm. (There is evidence that even when you only care about accuracy, private training sti...
eps和delta; 2.梯度裁剪,例如l2_norm_clip。注意的是2并不是原始dp sgd的约束。进一步...