Channel-masked Asymmetric Distribution Matching for Cross-Domain Generalized Dataset Distillation

Qi Liu, Chenghao Xu, Jiexi Yan, Guangtao Lyu, Erkun Yang, Guihai Chen, Yanhua YangAAAI 2026 · DOI

Abstract

Dataset distillation has achieved remarkable progress as an effective approach for data compression. However, real-world data often comes from diverse domains, leading to potential mismatches between the domains of synthesized images and those of the evaluation set. Existing methods primarily assume domain alignment between them, which limits their generalization ability in the above cross-domain scenarios. In this paper, we aim to ensure that images synthesized from known domains maintain robust performance on unseen domains and propose a novel framework called Channel-masked Asymmetric Distribution Matching (CADM). During asymmetric distribution matching, domain-sensitive channels of real data are selectively masked at different layers to extract domain-invariant features that guide synthetic data optimization. To further improve synthetic data representation, we introduce a class-focused domain-agnostic regularization to capture class-relevant knowledge while ignoring domain-specific information. Experiments show that our method produces domain-robust synthetic data and substantially improves generalization performance on unseen domains.

Introduction

Dataset Distillation (DD) aims to synthesize a small set of informative images that can effectively represent a large-scale dataset. While successful in controlled environments, standard DD methods often fail when the synthetic data and test data come from different distributions—a phenomenon known as "domain shift." For example, a model trained on distilled cartoon data may perform poorly on real-world photos. This paper identifies that standard distillation methods inadvertently capture domain-specific noise in certain feature channels, which hinders generalization to unseen domains.

The CADM Framework

To address this, the authors propose Channel-masked Asymmetric Distribution Matching (CADM). The framework focuses on identifying and masking "domain-sensitive" channels to force synthetic images to focus on universal, domain-invariant features.

The CADM framework consists of three main stages: (I) Pretraining with domain discriminators, (II) Asymmetric Distribution Matching via channel masking, and (III) Class-focused Domain-agnostic Regularization.

The methodology consists of three key components:

  1. Domain-Sensitive Channel Identification: During pre-training, domain discriminators are attached to network layers to identify feature channels that are most predictive of the source domain.
  2. Asymmetric Masking: During distillation, these domain-sensitive channels are masked in the real data. This forces the synthetic data to align its features with the remaining "content" channels, effectively filtering out stylistic quirks.
  3. Class-focused Domain-agnostic Regularization: A dual-consistency loss ensures the synthetic data relies on robust, class-relevant features. By masking synthetic data in different ways and requiring consistent classification results, the model learns to ignore fragile, domain-specific information.

Experimental Results

The researchers evaluated CADM across several multi-domain benchmarks, including PACS, VLCS, Office-Home, and DomainNet-Sub, using a "leave-one-domain-out" strategy.

Analysis of (a) performance degradation under domain shift, (b) channel robustness, and (c) the performance gains achieved by CADM over existing baselines.

CADM consistently outperformed state-of-the-art baselines. On the PACS dataset (ACS $\to$ Photo), CADM achieved 48.7% accuracy at 10 images per class (IPC), a significant improvement over the 40.8% achieved by the standard MMD baseline. Furthermore, synthetic data generated with CADM showed strong cross-architecture generalization, performing well even when transferred to different network architectures like VGG or AlexNet.

Conclusion

CADM provides a surgical approach to cross-domain dataset distillation by operating at the feature channel level. By masking domain-sensitive information during the matching process, it ensures that synthetic datasets capture robust, domain-invariant knowledge. While the method requires domain labels and adds some pre-training overhead, it represents a significant step toward making dataset distillation practical for real-world, multi-domain applications.

Summary updated

Abstract

Dataset distillation has achieved remarkable progress as an effective approach for data compression. However, real-world data often comes from diverse domains, leading to potential mismatches between the domains of synthesized images and those of the evaluation set. Existing methods primarily assume domain alignment between them, which limits their generalization ability in the above cross-domain scenarios. In this paper, we aim to ensure that images synthesized from known domains maintain robust performance on unseen domains and propose a novel framework called Channel-masked Asymmetric Distribution Matching (CADM). During asymmetric distribution matching, domain-sensitive channels of real data are selectively masked at different layers to extract domain-invariant features that guide synthetic data optimization. To further improve synthetic data representation, we introduce a class-focused domain-agnostic regularization to capture class-relevant knowledge while ignoring domain-specific information. Experiments show that our method produces domain-robust synthetic data and substantially improves generalization performance on unseen domains.
BibTeX
@inproceedings{liu2026channel,
  author = {Qi Liu and Chenghao Xu and Jiexi Yan and Guangtao Lyu and Erkun Yang and Guihai Chen and Yanhua Yang},
  title = {{Channel-masked Asymmetric Distribution Matching for Cross-Domain Generalized Dataset Distillation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2026},
  doi = {10.1609/aaai.v40i9.37661},
}
Figure