Scholarlink

<p>With the increasing focus on data privacy and security, Unsupervised Domain Adaptation (UDA) needs to be performed under a more stringent “source-free” setting, where only the model trained on the source data is available without accessing the source data itself. While domain alignment is a core concern in conventional UDA, this objective is not formulated in most existing Source-Free UDA (SF-UDA) approaches due to the difficulty of measuring domain discrepancy under the source-free scenario. Missing this mechanism, however, restricts their capability to effectively handle negative impact caused by domain shift. In this paper, we propose a new approach for source-free cross-domain knowledge transfer in aligned feature spaces. We exploit the statistical information of source training data stored in the Batch Normalization (BN) layers of the target model to construct the source domain distribution in feature spaces. The distribution discrepancy between the two domains can then be measured and minimized like conventional UDA without source data. In addition to the explicit formulation of domain alignment, we also apply self-supervised learning to discover target domain knowledge with unlabeled target data to enhance the encoder’s feature extraction capability. We conducted domain adaptation experiments on several standard image datasets. Results show that our method further improves the accuracy of existing source-free domain adaptation methods in various cases.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Source-free domain adaptation with aligned transfer and self-supervised learning

  • 期刊论文

  • 发表日期：

    2025年7月19日

查看全文
  • Yetao Weng,
  • Jian-Ping Mei,
  • Chengyang Hu,
  • Hongbo Zhao

摘要

With the increasing focus on data privacy and security, Unsupervised Domain Adaptation (UDA) needs to be performed under a more stringent “source-free” setting, where only the model trained on the source data is available without accessing the source data itself. While domain alignment is a core concern in conventional UDA, this objective is not formulated in most existing Source-Free UDA (SF-UDA) approaches due to the difficulty of measuring domain discrepancy under the source-free scenario. Missing this mechanism, however, restricts their capability to effectively handle negative impact caused by domain shift. In this paper, we propose a new approach for source-free cross-domain knowledge transfer in aligned feature spaces. We exploit the statistical information of source training data stored in the Batch Normalization (BN) layers of the target model to construct the source domain distribution in feature spaces. The distribution discrepancy between the two domains can then be measured and minimized like conventional UDA without source data. In addition to the explicit formulation of domain alignment, we also apply self-supervised learning to discover target domain knowledge with unlabeled target data to enhance the encoder’s feature extraction capability. We conducted domain adaptation experiments on several standard image datasets. Results show that our method further improves the accuracy of existing source-free domain adaptation methods in various cases.