Robust Source-Free Domain Adaptation From Non-Robust Source Models

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Unsupervised Domain Adaptation (UDA) often relies on robust source models, which are impractical.
• Source data may be inaccessible or inefficient for adaptation training in real-world scenarios.
• Existing methods struggle with adversarial training in UDA, leading to model degradation.

Purpose of the Study:

• To address robust source-free domain adaptation using only a non-robust source model and unlabeled target data.
• To develop a method that overcomes the degradation caused by adversarial training in UDA.
• To improve model robustness and performance in challenging domain adaptation tasks.

Main Methods:

• Proposed Source-Free Alternating Optimization (SFAO) to train a robust target model using a non-robust source model.
• Employed an alternating training strategy to minimize discrepancies between source and adversarial target domains.
• Introduced Softly-Constrained Adversarial Training (SCAT) to mitigate pseudo-label errors during adversarial training.

Main Results:

• SFAO significantly improves model performance on both clean and adversarial data.
• The proposed methods effectively address the challenges of robust source-free domain adaptation.
• Empirical findings show adversarial training amplification of UDA errors is mitigated.

Conclusions:

• Robust source-free domain adaptation is achievable with the proposed SFAO and SCAT methods.
• The approach offers a practical solution for scenarios lacking robust source models or source data.
• The study demonstrates a significant advancement in adversarial robustness for domain adaptation.