DGAF-Net Research Portal Deepfake Detection in MIS

Detailed Research Report

Evidence, workflow, architecture, and benchmark context.

This page expands the portal into a report-style presentation. It keeps the narrative strictly aligned with the published paper and the local repository assets, while surfacing the figures and derived plot set needed for a stronger public research package.

Research question

Can a hybrid architecture that combines DenseNet semantic learning with multi-scale texture-aware attention improve deepfake detection for compressed media relevant to MIS environments?

Compared models

The paper reports DGAF-Net against ViT-B16 and a Swin Transformer plus CNN fusion baseline under a shared training configuration.

Artifacts preserved here

Published PDF, LaTeX source, converted figure assets, original evaluation images, and static plots derived from the reported metric table and dataset summary.

Methodological Backbone

Balanced dataset setup and hybrid architecture rationale.

The report keeps the methodological story concise: dataset selection, preprocessing, baseline framing, and DGAF-Net’s semantic-texture fusion logic.

Workflow reference

Workflow overview

Data collection, preprocessing, model comparison, and evaluation are laid out in the original paper figure and reused here as a central navigation visual.

Method summary

  • 400-sequence FaceForensics++ subset from Kaggle, balanced between real and fake media.
  • 1 fps frame extraction, ImageNet normalization, and resizing for CNN and transformer inputs.
  • DenseNet121 semantic branch paired with a Gabor-driven multi-scale attention branch.
  • Gated attention fusion used to weight semantic and texture cues before classification.

Model Deep Dive

DGAF-Net is built around complementary semantic and forensic views.

DenseNet121 encourages feature reuse and semantic abstraction, while the Gabor path aims to surface directional texture anomalies that compressed deepfakes can still expose.

Architecture

DGAF-Net architecture

The original diagram highlights the DenseNet block, multi-scale attention block, Gabor filter stage, and the final classification head.

Design interpretation

  • Semantic branch: DenseNet121 feature extraction from facial image structure.
  • Texture branch: orientation-sensitive analysis through Gabor filtering.
  • Fusion: learned weighting through the gated attention fusion module.
  • Output: binary real-versus-fake classification tuned for digital-forensics relevance.

Reported Evaluation

DGAF-Net leads the benchmark comparison presented in the paper.

The table and figures below are grounded in the local paper source. No extra metrics have been added beyond what the project files support.

Class-wise performance table

Model Accuracy Class Precision Recall F1-score
ViT-B16 83.69% Real 84% 85% 84%
Fake 83% 82% 83%
Swin + CNN 85.00% Real 83% 91% 87%
Fake 88% 78% 82%
DGAF-Net 88.33% Real 88% 91% 90%
Fake 88% 85% 87%

Class-wise bar summary

Class-wise performance figure

Accuracy comparison

Accuracy comparison figure

Training and validation curves

Training and validation curves

ROC curves

ROC comparison figure

Confusion Analysis and Derived Plots

Original confusion images plus repository-native static analytics.

The repo preserves the bundled confusion-matrix images, while the new static plots are generated from the reported metrics to stay reproducible and GitHub-friendly.

ViT-B16 confusion matrix

ViT-B16 confusion matrix

Swin + CNN confusion matrix

Swin and CNN confusion matrix

DGAF-Net confusion matrix

DGAF-Net confusion matrix
Repository note

The paper narrative reports larger DGAF-Net confusion counts than the bundled confusion image. The derived plot assets therefore use the class-wise metric table and dataset description, which are internally consistent across the local project files.

DistPlot

Distplot of F1 scores

Pie Chart

Pie chart of dataset composition

Violin Plot

Violin plot of metric distributions

Heatmap

Heatmap of performance metrics

Pair Plot

Pair plot of class-wise metrics

Joint Plot

Joint plot of precision and recall

Reproducibility Notes

What this repository preserves and how the public-facing assets are generated.

The project’s original local directory was a compact paper archive. The repository build adds a publication-friendly wrapper around those assets without fabricating unsupported experimental detail.

Preserved source material

  • Published PDF preserved in both paper/published and docs/assets/paper.
  • Original LaTeX source kept under paper/latex-source.
  • Bundled figures copied into web-ready paths under docs/assets/images/figures.

Generated asset layer

  • scripts/generate_research_assets.py renders the PDF figures with PyMuPDF.
  • The same script emits static SVG plots from the paper’s reported metrics.
  • docs/assets/data/project-metrics.json records the evidence base used by the plots.