Progressive Debiasing and Boundary Adaptation for Source-Free Domain Adaptation

Abstract

Source-Free Domain Adaptation (SFDA) aims to adapt a pre-trained source model to an unlabeled target domain without accessing source data, offering a practical solution under strict privacy constraints. However, this paradigm exposes two tightly coupled challenges: persistent source-domain bias embedded within the model parameters and a self-reinforcing noisy pseudo-label loop during target adaptation. These factors jointly degrade cross-domain generalization and destabilize the adaptation process. To address this dilemma, we propose a novel Prune-to-Adapt (P2A) framework that resolves these challenges in a unified and progressive manner. Our core insight is that effective SFDA requires a two-stage strategy: first restoring model plasticity by alleviating entrenched source bias, and subsequently refining decision boundaries under noise-robust supervision. Accordingly, we introduce Bias-Alleviated Pruning (BAP), a structural debiasing mechanism that weakens source-specific parameter dependencies to yield a neutral initialization for adaptation. Building upon this debiased model, we develop Boundary-refined Negative Learning (BNL), which exploits high-confidence negative pseudo-labels to construct contrastive constraints. BNL effectively suppresses positive pseudo-label noise and explicitly sharpens class boundaries without requiring ground-truth labels. Through this progressive process of bias elimination and boundary refinement, P2A enables a systematic transition from a biased source model to a purified, target-adaptive model. Extensive experiments on challenging benchmarks, including Office-Home, DomainNet, and Office-31, demonstrate that our approach consistently outperforms state-of-the-art SFDA methods, achieving superior and robust cross-domain generalization.

Keywords

Source-Free Domain Adaptation / Model Pruning / Negative Learning / Pseudo-label Calibration

RIGHTS & PERMISSIONS

Higher Education Press 2026

Just Accepted

This article has successfully passed peer review and final editorial review, and will soon enter typesetting, proofreading and other publishing processes. The currently displayed version is the accepted final manuscript. The officially published version will be updated with format, DOI and citation information upon launch. We recommend that you pay attention to subsequent journal notifications and preferentially cite the officially published version. Thank you for your support and cooperation.