UAP (WACV 2026 oral)

Code for SSFL and UAP

Directory Structure

UAP/
├── train.py              # Main training script
├── evaluate.py           # Evaluation script
├── config.py             # Unified configuration module
│
├── methods/              # Model implementations
│   ├── SSFL.py          # Semi-supervised federated learning baseline
│   ├── UAP.py           # UAP method
│   └── base.py          # Base model class
│
├── datasets.py           # Dataset classes (PACS, VLCS, OfficeHome, etc.)
├── multi_client.py       # Multi-client data loading
├── utils.py              # Utility functions (model loading, aggregation, etc.)
├── losses/               # Loss functions (CDD, MMD, covariance)
├── mnist_datasets.py     # MNIST dataset utilities
│
├── requirements.txt      # Python dependencies
└── BATCHNORM_STATS.md   # Documentation on BN stats sharing

Quick Start

1. Install Dependencies

conda create -n uap python=3.8
conda activate uap
pip install -r requirements.txt

2. Training

# Train SSFL on PACS
python train.py --dataset PACS --test_env 0 --method SSFL --device cuda:0

# Train UAP on PACS
python train.py --dataset PACS --test_env 0 --method UAP --device cuda:0

3. Evaluation

# Evaluate SSFL
python evaluate.py --dataset PACS --test_env 0 --method SSFL --experiment_path data_final --device cuda:0

# Evaluate UAP
python evaluate.py --dataset PACS --test_env 0 --method UAP --experiment_path data_final --device cuda:0

Configuration

Recommended

from config import get_args
args = get_args()

Important Arguments

• --dataset: Dataset name (PACS, VLCS, OfficeHome)
• --test_env: Target test domain (0-3 for PACS)
• --server_domain: Server labeled domain (0-3)
• --method: Method name (SSFL, UAP)
• --multi_client: Enable multi-client setting
• --num_clients: Number of clients in multi-client mode
• --rounds: Number of federated rounds
• --E: Local epochs per round
• --device: CUDA device (cuda:0, cuda:1, cpu)

Datasets

Required datasets:

• PACS: 4 domains (art_painting, cartoon, photo, sketch)
• VLCS: 4 domains (Caltech101, LabelMe, SUN09, VOC2007)
• OfficeHome: 4 domains (Art, Clipart, Product, Real_World)

Place datasets in ./data/ directory:

data/
├── PACS/
│   ├── art_painting/
│   ├── cartoon/
│   ├── photo/
│   └── sketch/
├── VLCS/
│   ├── Caltech101/
│   ├── LabelMe/
│   ├── SUN09/
│   └── VOC2007/
└── officehome/
    ├── Art/
    ├── Clipart/
    ├── Product/
    └── Real_World/

Results

Checkpoints and logs are saved to:

results/{experiment_path}/{dataset}/{dataset}/{method}/target_{test_env}/server_{server_domain}/seed_{seed}/

Each checkpoint includes:

• checkpoint.pt: Model weights and optimizer state
• target_accs.csv: Target domain accuracies per round

Documentation

• BATCHNORM_STATS.md: Explains batch normalization statistics sharing and privacy preservation

Module Dependencies

train.py

train.py
  ├── config.py (or argument.py)
  ├── utils.py
  │   └── methods/ (SSFL, UAP)
  │       └── base.py
  ├── datasets.py
  └── multi_client.py

evaluate.py

evaluate.py
  ├── prepare.py (or config.py)
  ├── utils.py
  ├── datasets.py
  └── losses/