Research map / Label-efficient and robust learning / Domain adaptation and generalization

Vicinity-Guided Discriminative Latent Diffusion for Privacy-Preserving Domain Adaptation

NeurIPS 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Conditional distribution alignment · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Discriminative Vicinity Diffusion (DVD), a drift-only latent diffusion module shared with the source-free adaptation setting, which generates source-like latent cues from target k-NN priors to align a target encoder via InfoNCE.

Problem

Source-free domain adaptation lacks explicit knowledge transfer from the source model; the paper asks whether latent diffusion models can enable explicit discriminative transfer when the source provider can share an auxiliary module but not raw source data.

Approach

Encodes each source sample's label information into a Gaussian prior over its latent k-NN vicinity and trains a deterministic drift-only diffusion network to map vicinity samples back to label-consistent features; at adaptation, the frozen module generates source-like features from target vicinities, which guide a contrastive alignment of the target encoder.

Key results

• On VisDA-C 2017 (ResNet-101), DVD reaches 88.9% average accuracy, versus 87.8% for the best listed prior method (NRC++).
• On Office-Home (ResNet-50), DVD reaches 73.7% average accuracy, above DaC (72.8%) and CPD (73.0%).
• On CIFAR-10 with ResNet-18, DVD lowers top-1 error from 7.07% to 6.52%.

Measured effects reported in the paper

Surprising findings

• Adding Gaussian noise at the input or latent level degrades target accuracy relative to the baseline on VisDA-C, while the k-NN vicinity prior improves it.
• Stochastic (noise-injected) drift reduces accuracy by 1.7% compared with deterministic drift on VisDA-2017.

Limitations

• Benchmark gains are modest on several settings (e.g., DVD underperforms CPD-style baselines on some Office-31 and VisDA classes, and is below ZS-CLIP-based PromptStyler on VLCS), and reported std values are small relative to gaps only in some cases.
• The privacy-preserving claim rests on shipping a diffusion module trained on source features; no formal privacy analysis is given.
• Several appendix comparisons (e.g., Table on VisDA stochastic vs deterministic) are described with limited detail.

Experimental setup

Datasets: Office-31 Office-Home VisDA-C 2017 CIFAR-10 CIFAR-100 ImageNet PACS VLCS DomainNet

Benchmarks: VisDA-C 2017 Office-Home Office-31 CIFAR-10 CIFAR-100 ImageNet PACS VLCS DomainNet

Metrics: classification accuracy top-1 error runtime per epoch inference time FPS

Baselines: SHOT HCL DaC NRC++ SFADA CPD 3C-GAN AaD ZS-CLIP CAD PromptStyler

Builds on: SHOT NRC InfoNCE SimCLR flow matching

Closest papers by meaning

• Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation (NeurIPS 2025 · similarity 0.89)
• A General Representation-Based Approach to Multi-Source Domain Adaptation (ICML 2025 · similarity 0.88)
• Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.88)
• Unsupervised Region-Based Image Editing of Denoising Diffusion Models (AAAI 2025 · similarity 0.87)
• Domain Guidance: A Simple Transfer Approach for a Pre-trained Diffusion Model (ICLR 2025 · similarity 0.87)
• Transfer Learning for Diffusion Models (NeurIPS 2024 · similarity 0.87)