Algorithm 2: CMD Module
[1]	Input: $F_{noisy}$, $S_{clean}$
	Parameters: $W_q,W_k,W_v$
[2]	Output: Denoised features $F_{denoised}$
1:	procedure FORWARD ($F_{noisy}$, $S_{clean}$)
2:	Common latent space
3:	$\tilde{F} =F_{noisy}{W}_F$, $\tilde{S} =S_{clean}{W}_S$
4:	Noise-aware queries
5:	$Q^{\prime }={\displaystyle \frac{\tilde{F}{W}_Q}{\sqrt{d}\hspace{0.17em}(1+\widehat{\sigma })}}$                                    # Query from noisier modality
6:	Key and value metrics
7:	$K =\tilde{S}{W}_K$                                               # Key from clean modality
8:	$V =\tilde{S}{W}_V$                                            # Value from clean modality
9:	Scaled dot-product
10:	$A =Softmax\left(Q^{\prime }{K}^T\right)V$
11:	Semantic gating mask
12:	$G =\mathit{\sigma }(W_g\tilde{S}\hspace{0.17em})$
13:	Amplified Attention
14:	$\widehat{F}=A\odot G$
15:	Residual fusion
16:	$F^{*}=\tilde{F}+\widehat{F}$
17:	GDFN
18:	return $\mathit{LN}({\tilde{F}}_{noisy} + [\varphi \left(W_1^{\left(d\right)}{W}_1^{\left(p\right)}{\tilde{F}}_{noisy}\right)\odot \left(W_2^{\left(d\right)}{W}_2^{\left(p\right)}{\tilde{F}}_{noisy}\right)\left]\right)$
19:	end procedure