Robust Unsupervised Domain Adaptation from A Corrupted Source

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Unsupervised Domain Adaptation (UDA) enables learning without labeled data by transferring knowledge from source to target domains.
• Current UDA methods are vulnerable to corrupted source data, including inherent noise and adversarial attacks.
• This fragility limits the practical application of UDA in real-world scenarios.

Purpose of the Study:

• To develop a robust framework for Unsupervised Domain Adaptation (UDA) that effectively handles corrupted source domain data.
• To enhance the resilience of UDA models against both inherent data corruption and adversarial poisoning attacks.
• To achieve high performance on target domains even when source data quality is compromised.

Main Methods:

• Proposes a novel framework for UDA from corrupted source domains.
• Employs knowledge ensemble by learning multiple domain-invariant models on random data partitions.
• Refines models using mutual information maximization to adaptively capture high-confidence predictive information from the target domain.

Main Results:

• Demonstrates robustness against various types of poisoned data attacks.
• Achieves high asymptotic performance on the target domain despite source data corruption.
• The proposed approach effectively addresses distribution shifts in corrupted UDA settings.

Conclusions:

• The developed framework offers a principled and effective solution for UDA with corrupted source data.
• The combination of knowledge ensemble and mutual information maximization enhances model adaptability and robustness.
• This work advances the field of UDA by enabling reliable knowledge transfer from noisy or attacked source domains.