Algorithm 1. Framework of proposed P2P

Require: previous model ${\mathcal{M}}_{t-1}$ , a batch dataset $D={\left\{x_i,y_i\right\}}_{i=1}^n$ , the probability threshold p, the number of iterators k, learning rate $\alpha$ .

Ensure: A translated dataset $D_{trans}$ , a private domain dataset $D_{pri}$ , a public domain dataset $D_{pub}$ .

1. Freeze ${\mathcal{M}}_{t-1}$

2. Compute the predicted probability distribution for $D:P=\text{Softmax}\left({\mathcal{M}}_{t-1}\left(x\right)\right)$

3. Divide $D:D_{pub}=\left\{\left(x_i,y_i\right)|P_{y_i}\ge p\right\},D_{trans}=\left\{\left(x_i,y_i\right)|P_{y_i}<p\right\}$

4. For $x\in D_{pri}$ , initialize $x^{*}=x+\delta$ with a small random noise $\delta$

5. for $iter=\text{1}$ to k do

6. $\delta ={\nabla }_{x^{*}}\left[{\mathcal{L}}_{\text{cross-entropy}}\left({\theta }_{t-1};x^{*},y\right)\right]$

7. $x^{*}=x^{*}-\alpha \frac{1}{{\Vert \delta \Vert }_2}$

8. end for

9. $D_{trans}=\left\{\left(x^{*},y\right)|\overline{y}\left(x^{*},{\theta }_{t-1}\right)=y\right\}$

10. Return $D_{trans}$ , $D_{pri}$ , $D_{pub}$