Robust classification in noisy environments remains a fundamental challenge in machine learning. Standard approaches typically treat signal enhancement and classification as separate, sequential stages: first enhancing the signal and then applying a classifier. This approach fails to leverage the semantic information in the classifier's output during denoising. In this work, we propose a general, domain-agnostic framework that integrates two interacting diffusion models: one operating on the input signal and the other on the classifier's output logits, without requiring any retraining or fine-tuning of the classifier. This coupled formulation enables mutual guidance, where the enhancing signal refines the class estimation and, conversely, the evolving class logits guide the signal reconstruction towards discriminative regions of the manifold. We introduce three strategies to effectively model the joint distribution of the input and the logit. We evaluated our joint enhancement method for image classification and automatic speech recognition. The proposed framework surpasses traditional sequential enhancement baselines, delivering robust and flexible improvements in classification accuracy under diverse noise conditions.
Figures & tables
Figure 1: Comparison of noisy signal classification paradigms. xcor , ycor denote the corrupted signal and logits; x^0 , y^0 their denoised versions. (a) Noisy: direct classification. (b) Enhanced: independent signal enhancement followed by classification. (c) CARD: denoising logits, y , given fixed corrupted signal, xcor . (d) Proposed: coupled denoising of both signal and logits.
Figure 2: Overview of the proposed architectures. (a) Parallel : x and y are denoised simultaneously, exchanging estimates x^0 and y^0 at each timestep t for step-wise cross-conditioning. (b) Alternating : an iterative refinement strategy where complete diffusion trajectories for x and y condition each other in alternating turns. (c) Nested : a guidance mechanism where every single step of the outer y process triggers a full inner diffusion loop over x .
Figure 3: Qualitative comparison under 30% GN. The enhanced baseline incorrectly recognizes “2,” whereas our coupled strategies ( Alternating , Parallel , and Nested ) correctly recover the digit by leveraging logit-space guidance, reconstructing the missing upper-left stroke, enabling correct recognition as “8.”
Baselines (%) ↑
Coupled (Ours) (%) ↑
Dataset
Corruption
Noisy
Enhanced
URIE
CARD
Parallel
Nested
Alternating
MNIST
15% GN
89.7
98.4
99.4
86.3
99.2
99.3
98.9
30% GN
66.7
98.3
99.3
66.2
99.2
99.1
98.7
Gaussian Blur
89.6
98.3
99.5
87.0
99.2
99.3
98.6
CIFAR-10
15% GN
16.6
60.9
68.5
11.3
74.3
81.3
83.0
30% GN
10.9
60.4
59.4
10.4
71.6
81.2
82.8
Table 1: Classification accuracy (%) on image datasets, with pre-trained classifier (Noisy), Enhanced signal baseline, URIE (enhancer trained with regression and classification losses), static conditional generation of logits (CARD), and our coupled strategies. Best robust results are highlighted in bold.
Baselines (%) ↓
Coupled (Ours) (%) ↓
Dataset
Noisy
Enhanced
Dissen
CARD
Parallel
Nested
Alternating
G.Cmds-Reverb
5.58
5.11
7.96
11.08
4.98
4.85
5.03
EARS-Reverb
5.79
4.48
6.46
16.79
3.95
3.83
4.25
EARS-WHAM
14.02
10.04
13.04
36.58
9.90
9.75
10.16
Table 2: WER (%) comparison on Google Commands, EARS-Reverb and EARS-WHAM. Noisy denotes the performance of Whisper-base. Enhanced denotes audio-only enhancement, Dissen denotes an enhancer trained with regression and classification losses, and CARD denotes logits-only enhancement.
Figure 4: Robustness vs. Sampling Steps. Accuracy as a function of the number of diffusion steps on CIFAR-100 and ImageNet32-100 under 30% GN. The Parallel strategy converges rapidly, while Nested and Alternating require more steps to saturate.
Dataset
Corruption
Clean-estimate guidance
Noisy-sample guidance
CIFAR-100
15% GN
70.7
70.0
30% GN
68.5
67.0
Gaussian Blur
64.8
64.0
IN32-100
15% GN
69.8
69.6
30% GN
65.5
64.6
Gaussian Blur
63.3
61.8
Table 3: Effect of guidance source in the Parallel strategy. We compared conditioning on the clean estimate ( x^0(t) , y^0(t+1) ) versus the noisy sample ( xt−1 , yt ). Classification accuracy (%) is reported.
Corruption
DDPM
DDIM
15% GN
73.5
70.7
30% GN
71.2
68.5
Gaussian Blur
69.7
64.8
Table 4: Evaluation of DDPM versus DDIM sampling algorithms for the Parallel strategy on CIFAR-100. Classification accuracy (%) is reported.
Corruption
Enhanced
Parallel (Ours)
15% GN
92.1
94.0
30% GN
86.5
92.4
Gaussian Blur
94.0
94.1
Table 5: Scaling to full-resolution ImageNet ( 224×224 , 10 classes). Classification accuracy (%) under different corruptions. We compare the Enhanced baseline with our Parallel strategy.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Qualitative signal enhancement examples on MNIST. Each row corresponds to a different test sample. From left to right: corrupted input (ground truth shown for reference), Enhanced baseline, Alternating, Parallel, and Nested strategies. The predicted class under each reconstruction is indicated in the figure. Our coupled strategies consistently recover digit structures that are visually clearer and semantically aligned with the true class, leading to correct predictions.
Shenzhen Research Institute, The Chinese University of Hong Kong · University of Science and Technology of China · Imperial Global Singapore, Imperial College London