A3: Active Adversarial Alignment for Source-Free Domain Adaptation

TL;DR

A3 addresses source-free unsupervised domain adaptation by integrating SwAV-inspired self-supervised pretraining, an active sampling strategy that jointly optimizes uncertainty and diversity, and adversarial plus regularization losses to align target representations with the (unavailable) source distribution. The method constructs a core-set from targeted samples via a Bayesian acquisition framework, and enforces domain-invariant features through a domain adversary and local perturbation-based regularization. Key contributions include a hybrid acquisition function, a novel combination of DAL and VAT losses, and extensive ablations demonstrating the value of joint active sampling and adversarial alignment. Empirically, A3 achieves state-of-the-art performance on Office-31, Office-Home, and DomainNet benchmarks, maximizing robustness to label noise and distribution shift in SFUDA scenarios.

Abstract

Unsupervised domain adaptation (UDA) aims to transfer knowledge from a labeled source domain to an unlabeled target domain. Recent works have focused on source-free UDA, where only target data is available. This is challenging as models rely on noisy pseudo-labels and struggle with distribution shifts. We propose Active Adversarial Alignment (A3), a novel framework combining self-supervised learning, adversarial training, and active learning for robust source-free UDA. A3 actively samples informative and diverse data using an acquisition function for training. It adapts models via adversarial losses and consistency regularization, aligning distributions without source data access. A3 advances source-free UDA through its synergistic integration of active and adversarial learning for effective domain alignment and noise reduction.

Figures (2)

• Figure 1: Our framework involves two main phases: the source model pretraining and the model adaptation. We begin by adopting a training regime for the source model, which is initialized with a pre-trained ImageNet model deng2009imagenet. This is referred to as stage-0 in the multi-stage adaptation process. The second phase encompasses n-1 stages of active learning. During this phase, a core-set is constructed and the Bayesian model uncertainty is estimated. At each iteration, the core-set is updated with the top-k informative samples from the training pool. These samples are then used to retrain the Bayesian and target models until the data sampling budget is exhausted. Refer to Section \ref{['sec:proposed_method']} for further details on target domain alignment. Za and Zb denote the two augmentations of the input used for self-supervised training of the source and target models.
• Figure 2: T-SNE plot of the learned source features and the target features at each active learning adaptation cycle.