Noise Optimized Conditional Diffusion for Domain Adaptation

TL;DR

NOCDDA addresses unsupervised domain adaptation when high-confidence pseudo-labels in the target domain are scarce, which harms cross-domain alignment. It unifies forward diffusion-based classifier training with a conditional diffusion generator and introduces class-aware terminal distributions to guide reverse diffusion, producing discriminative, hcpl-tds-aligned target samples. Key contributions include time-embedding aware classifier unification, conditional adversarial domain alignment, forward diffusion with cross-domain consistency, and class-specific noise-optimized backward sampling, all validated across 5 datasets and 29 tasks with strong gains over 31 baselines. The approach demonstrates robust cross-domain decision boundaries and effective target-domain representation, offering significant practical impact for real-world UDA scenarios with limited hcpl-tds.

Abstract

Pseudo-labeling is a cornerstone of Unsupervised Domain Adaptation (UDA), yet the scarcity of High-Confidence Pseudo-Labeled Target Domain Samples (\textbf{hcpl-tds}) often leads to inaccurate cross-domain statistical alignment, causing DA failures. To address this challenge, we propose \textbf{N}oise \textbf{O}ptimized \textbf{C}onditional \textbf{D}iffusion for \textbf{D}omain \textbf{A}daptation (\textbf{NOCDDA}), which seamlessly integrates the generative capabilities of conditional diffusion models with the decision-making requirements of DA to achieve task-coupled optimization for efficient adaptation. For robust cross-domain consistency, we modify the DA classifier to align with the conditional diffusion classifier within a unified optimization framework, enabling forward training on noise-varying cross-domain samples. Furthermore, we argue that the conventional \( \mathcal{N}(\mathbf{0}, \mathbf{I}) \) initialization in diffusion models often generates class-confused hcpl-tds, compromising discriminative DA. To resolve this, we introduce a class-aware noise optimization strategy that refines sampling regions for reverse class-specific hcpl-tds generation, effectively enhancing cross-domain alignment. Extensive experiments across 5 benchmark datasets and 29 DA tasks demonstrate significant performance gains of \textbf{NOCDDA} over 31 state-of-the-art methods, validating its robustness and effectiveness.

Figures (4)

• Figure 1: Illustration of the proposed NOCDDA method. Fig.\ref{['fig:1']}(a) shows source (red) and target (blue) data with distinct distributions and geometric structures, where shapes indicate class labels. Fig.\ref{['fig:1']}(b&c) depicts generator training on labeled source data and pseudo-label prediction for the target domain. Fig.\ref{['fig:1']}(d) aligns feature and label spaces via conditional adversarial learning. hcpl-tds selection is highlighted in Fig.\ref{['fig:1']}(e), followed by robust cross-domain classifier training under forward noise perturbation in Fig.\ref{['fig:1']}(f). Fig.\ref{['fig:1']}(g&h) illustrates class-specific backward sampling and enriched target sample generation. Fig.\ref{['fig:1']}(i) shows enhanced DA performance via improved target representation.
• Figure 2: Accuracy${\rm{\% }}$ on Digital Image Datasets.
• Figure 3: Accuracy${\rm{\% }}$ on Office-31 Datasets.
• Figure 4: Ablation Study on Digital Image Datasets.