Algorithm 1. Procedure of the proposed framework.
	Require: labeled dataset ${\mathcal{D}}^{\mathcal{L}}$, unlabeled dataset ${\mathcal{D}}^{\mathcal{U}}$
	Ensure: classifier $\mathcal{M}$
1:	Construct and initialize class classifier $\mathcal{M}$, encoder $E$, generator $G$, real/fake classifier $D$
2:	Train $\mathcal{M}$, $E$, $D$, $G$ with ${\mathcal{D}}^{\mathcal{L}}$
3:	while not satisfy condition do
4:	randomly sample $x_{batch}$ from ${\mathcal{D}}^{\mathcal{U}}$
5:	calculate $scor{e}_i$ $\forall$ $i$ $\in$ $x_{batch}$ according to (13)
6:	select $n$ sample with the highest score
7:	$D_{inc}\leftarrow \varnothing$
8:	for $x^{*}$ in $x_{re}$ do
9:	query the label of $x^{*}$ from Oracle and get $y^{*}$ $x_{re}^{*}\leftarrow G\left(E\left(x^{*}\right)\right)$
10:	$x_{re}^{*}\leftarrow G\left(E\left(x^{*}\right)\right)$
11:	$D_{inc}\leftarrow D_{inc}\cup \left\{\left(x_{re}^{*},y^{*}\right),\left( x^{*}, y^{*}\right)\right\}$
12:	${\mathcal{D}}^{\mathcal{L}}\leftarrow {\mathcal{D}}^{\mathcal{L}}\cup D_{inc}$
13:	end for
14:	retrain $\mathcal{M},E,D,C$ with ${\mathcal{D}}^{\mathcal{L}}$ by (2)–(8)
15:	return $\mathcal{M}$