DURA

DURA: Dual-Stream LoRA with Reliability-Aware Fusion for Multi-Domain Multimodal Fake News Detection

DURA is a multimodal fake news detection framework designed for domain-aware multi-domain learning under domain discrepancy and modality irregularity. It encourages separation between shared forensic priors with reduced domain specificity and domain-sensitive residual semantics, then uses reliability-aware visual aggregation and gated multimodal fusion to reduce the influence of weak, missing, or noisy visual evidence.

Framework

Overview

Multimodal fake news on social media often contains heterogeneous domain styles, imbalanced domain distributions, noisy multi-image evidence, and incomplete modalities. DURA addresses these issues with a dual-stream LoRA architecture:

• A frozen shared stream models shared forensic priors with reduced domain specificity.
• A trainable private stream captures domain-sensitive residual semantics.
• A semantic orthogonality constraint reduces redundant overlap between the two streams.
• DARE-based adapter merging constructs the shared stream from domain-specific LoRA adapters.
• Reliability aggregation estimates task-relevant visual evidence and suppresses weak or noisy image instances.
• Gated fusion dynamically recalibrates text-image contributions for the final prediction.

Repository Structure

dura/
  config.py                  Configuration dataclasses and YAML loading
  data.py                    Dataset construction and batch preparation
  feature_cache.py           Feature caching utilities
  model.py                   DURA model and backbone wrappers
  pipeline.py                End-to-end experiment pipeline
  train.py                   Training entry definitions
  common/                    Logging, checkpointing, seeding, and I/O helpers
  data_utils/                Data loading, cleaning, sampling, and split helpers
  eval_utils/                Metrics, reports, and threshold utilities
  modules/                   LoRA adapters and loss functions
  pipeline_utils/            DARE adapter merging utilities
  viz/                       Representation visualization utilities

scripts/
  run_dura.py                Command-line entry for training and evaluation
  run_dura_robustness_eval.py
                              Missing-modality and visual-conflict diagnostics
  summarize_dura_robustness.py
                              Summarize robustness runs into mean/std tables

assets/
  figure2.pdf                Framework diagram in PDF format
  figure2.png                Framework diagram preview for GitHub rendering

requirements.txt             Python dependency list
LICENSE                      Apache-2.0 license

Method Highlights

Dual-stream feature disentanglement. DURA uses separate shared and private LoRA streams to model shared fake-news detection cues with reduced domain specificity and domain-sensitive residual cues.

DARE-based shared stream construction. Domain-specific LoRA adapters are merged into a shared plugin through DARE-style sparsification and rescaling, which helps preserve useful common knowledge while reducing domain interference.

Orthogonality-constrained representation learning. A semantic orthogonality loss encourages the shared and private post-level representations to capture complementary information.

Reliability aggregation for visual evidence. Multi-image posts are handled with reliability weighting so that informative visual instances receive stronger aggregation weights than weakly relevant or noisy instances.

Gated multimodal fusion. DURA uses domain reintegration gates and a modal gate to recalibrate fused text-image representations before classification.

The reliability factors in this implementation should be interpreted as task-related evidence-quality weights rather than calibrated confidence values for autonomous moderation decisions.

Installation

git clone https://github.com/ALateFall/DURA.git
cd DURA

python -m venv .venv
source .venv/bin/activate

pip install --upgrade pip
pip install -r requirements.txt

Install the PyTorch build that matches your CUDA environment if the default package resolver does not select the desired CUDA version.

Data Preparation

For the Weibo and Weibo21 datasets, please request access from the original dataset authors. For GossipCop, please obtain the benchmark data from the official FakeNewsNet release. After obtaining the benchmark data, prepare a local YAML configuration file that points to your dataset location.

This repository intentionally does not release datasets, local YAML configuration files, cached features, checkpoints, run outputs, or train/val/test split maps. These files are environment- and permission-dependent and should be kept outside version control.

Training and Evaluation

DURA is configured through a YAML file whose fields correspond to the dataclasses in dura/config.py. The command below runs training, validation threshold selection, and final test evaluation:

python scripts/run_dura.py --config path/to/dura_config.yaml

Configuration values can also be overridden from the command line:

python scripts/run_dura.py --config path/to/dura_config.yaml \
  --override paths.dataset_root=/path/to/authorized_dataset \
  --override paths.output_root=outputs/dura_run \
  --override data.dataset_name=weibo21 \
  --override clip.local_files_only=false

Component-removal variants can be launched with the same entry point:

python scripts/run_dura.py --config path/to/dura_config.yaml \
  --override experiment.ablation=w/o_ivw

python scripts/run_dura.py --config path/to/dura_config.yaml \
  --override experiment.ablation=w/o_adaptive

Robustness diagnostics can be evaluated from an existing DURA run directory:

python scripts/run_dura_robustness_eval.py \
  --source-run-dir outputs/example_run \
  --nested-noise \
  --donor-label-strategy opposite_label

The opposite-label donor strategy corresponds to the visual-conflict replacement stress test described in the paper. Missing-text robustness is implemented as evaluation-time zeroing of cached text features; it is not a separate learned text-mask module.

Outputs

Each run writes its artifacts under the configured paths.output_root, including:

<output_root>/<run_name>/
  run_config.json
  summary.json
  split_map.json
  audit/
  checkpoints/
    dura_best.pt
    val_report.json
    test_report.json
    test_per_domain.csv
    predictions_val.json
    predictions_test.json

Do not commit run outputs, split maps, local configuration files, cached features, or checkpoints.

Benchmarks

The method is evaluated in the paper on three representative multimodal fake news detection benchmarks:

• Weibo
• Weibo21
• GossipCop

Citation

If you use DURA in your research, please cite:

@article{pu2026dura,
  title  = {DURA: Dual-Stream LoRA with Reliability-Aware Fusion for Multi-Domain Multimodal Fake News Detection},
  author = {Pu, Ao and Feng, Xia and Liang, Gang and Zhao, Kui and Hu, Haixin and Xu, Yani and Wang, Lei},
  year   = {2026}
}

License

This project is released under the Apache License 2.0.