DGAF-Net Research Portal Deepfake Detection in MIS

IEEE COMPAS 2025 | Published October 23, 2025

Deepfake Detection in MIS

Leveraging DenseNet and multi-scale information for enhanced digital forensics.

DGAF-Net fuses DenseNet121 semantic features, Gabor-driven texture evidence, and gated attention fusion to strengthen deepfake detection under compression, noise, and other real-world distortions that matter in management information systems.

Open Research Report Secure Downloads GitHub Repository Watch Video
  • Balanced FaceForensics++ subset with 400 video sequences, sampled at 1 frame per second.
  • Hybrid comparison against ViT-B16 and Swin Transformer plus CNN fusion baselines.
  • IEEE-published evaluation focused on MIS trust, media integrity, and digital forensics.

Publication Snapshot

First page preview of the IEEE paper
FaceForensics++ DenseNet121 Gabor Filters Gated Attention Fusion Digital Forensics MIS
0% Best reported overall accuracy for DGAF-Net
0 ROC AUC reported for the proposed model
0 FaceForensics++ sequences in the evaluated subset
0 Training epochs under the shared optimization setup

Lead Researcher Profile

Md Anisur Rahman Chowdhury

Deepfake detection, digital forensics, MIS trust, applied AI, and research-to-public artifact development. This portal highlights the lead author and connects every primary public profile for authorship verification, collaboration, and citation discovery.

MAC
Primary author

Md Anisur Rahman Chowdhury

Featured researcher for the DGAF-Net deepfake-detection study and the public research package.

Open Portfolio Open GitHub
Code and artifacts

GitHub

Repositories, research portals, implementation assets, and source visibility.

github.com/ANIS151993
Portfolio

Portfolio Website

Personal project hub and extended research showcase.

marcbd.site

Why This Matters

Deepfakes are a data integrity problem, not only a media problem.

In MIS settings, manipulated visual evidence can distort decision support, incident response, reputational management, and trust in operational data pipelines. The project frames deepfake detection as a practical digital-forensics requirement for modern information systems.

Decision Integrity

Visual misinformation can influence executive judgment, investigations, and downstream analytics.

Compression Resilience

The paper emphasizes realistic compressed media, reflecting how forged content circulates on social platforms and enterprise channels.

Operational Relevance

The proposed architecture aims to remain efficient enough for practical digital-forensics workflows rather than only benchmark-driven evaluation.

Workflow and Model

From curated FaceForensics++ inputs to fused semantic-texture evidence.

The research pipeline combines dataset curation, preprocessing, baseline benchmarking, and the DGAF-Net architecture that merges DenseNet features with Gabor-based texture evidence through a gated attention fusion mechanism.

Project Workflow

Workflow overview from the paper

The paper workflow covers data collection, preprocessing, model comparison, and evaluation using accuracy, precision, recall, F1, and confusion analysis.

DGAF-Net Architecture

Architecture of the proposed DGAF-Net model

DenseNet121 contributes high-level semantic structure, while a multi-scale Gabor attention branch focuses on texture irregularities often associated with facial manipulation artifacts.

Dataset and Training

Evaluation grounded in a balanced, compressed FaceForensics++ subset.

Only the settings explicitly described in the paper are presented here. The portal avoids inventing unreleased implementation details or unsupported dataset claims.

Dataset profile

  • 400 video sequences: 200 real and 200 fake.
  • Manipulations: DeepFakes, Face2Face, FaceSwap, NeuralTextures.
  • Compression approximates C23 quality for realistic platform conditions.

Preprocessing

  • Frames extracted at 1 frame per second to reduce redundancy.
  • ImageNet mean and standard deviation normalization.
  • Input resizing and optional randomized JPEG degradation during preparation.

Training setup

  • Binary cross-entropy optimized with AdamW.
  • Learning rate: 1e-4 with cosine annealing.
  • Batch size 32, 20 epochs, dropout 0.3, and L2 weight decay.

Results Snapshot

DGAF-Net leads the reported benchmark comparison across accuracy and AUC.

The site preserves the paper’s exact reported comparison values for ViT-B16, Swin Transformer plus CNN fusion, and the proposed DGAF-Net model.

Performance Comparison

Model Accuracy Precision Recall F1-score AUC
ViT-B16 83.69% 84 / 83 85 / 82 84 / 83 0.84
Swin + CNN 85.00% 83 / 88 91 / 78 87 / 82 0.86
DGAF-Net 88.33% 88 / 88 91 / 85 90 / 87 0.88

Precision, recall, and F1 entries are shown as Real / Fake class values exactly as reported in the paper.

Accuracy Trend

Accuracy comparison chart

Training Dynamics

Training and validation curves

ROC Comparison

ROC curve comparison

Analytics Dashboard

Static plot assets derived from the paper’s reported evidence base.

These visuals were generated for this repository from the verified metric table and dataset summary, then exported as static web assets for both GitHub and GitHub Pages.

DistPlot

Distplot of reported F1 scores

Distribution of the six reported class-wise F1 observations.

Pie Chart

Pie chart of dataset composition

Balanced FaceForensics++ subset composition used by the study.

Violin Plot

Violin plot of model metric distributions

Per-model distribution of reported precision, recall, and F1 values.

Heatmap

Heatmap of performance matrix

Performance matrix across accuracy, AUC, and class-wise metrics.

Pair Plot

Pair plot of class-wise metrics

Precision, recall, and F1 relationships across model-class observations.

Joint Plot

Joint plot of precision vs recall

Precision versus recall with marginal histograms and a trend line.

Video Overview

Project walkthrough and research communication.

The project video extends the paper with a quick-access explainer suitable for reviewers, collaborators, and visitors arriving from the repository or scholar profile.

Publication and Authors

Connected to the paper, scholar profile, and supporting repository assets.

This repository package is designed as a durable research artifact: publication-aware, link-rich, and clear about what is sourced directly from the paper.

Suggested citation

@inproceedings{rahmanchowdhury2025deepfake,
  title     = {Deepfake Detection in MIS: Leveraging DenseNet and Multi-Scale Information for Enhanced Digital Forensics},
  author    = {Md Anisur Rahman Chowdhury and Khandakar Rabbi Ahmed and Shahriar Alam Robin and Shah Tawkir Nesar and Towfika Salam and Md. Naimul Ahsan and Md Eahia Ansari},
  booktitle = {2025 IEEE 2nd International Conference on Computing, Applications and Systems (COMPAS)},
  year      = {2025},
  month     = {October},
  day       = {23},
  url       = {https://ieeexplore.ieee.org/abstract/document/11381812}
}

Author list from the paper

Md Anisur Rahman Chowdhury, Khandakar Rabbi Ahmed, Shahriar Alam Robin, Shah Tawkir Nesar, Towfika Salam, Md. Naimul Ahsan, and Md Eahia Ansari.

The publication connects institutions in the United States and Bangladesh, reinforcing the project’s interdisciplinary positioning across digital forensics, MIS, and applied AI.

IEEE Xplore Google Scholar

Document Security Policy

Protected access flow for paper and source documents.

This portal follows the same public policy style as the follow-up GitHub research website: visitors first verify the author channels, then request access before opening the preserved document bundle.

Step 1

Follow and verify public channels

Open the lead researcher's GitHub profile and project video, then confirm the checks in the gate.

Open GitHub Profile Open Video
Step 2

Request document access

Send a short email request so the author can verify academic, review, or collaboration purpose before sharing protected materials.

Email the Author
Step 3

Open secure download gate

Complete the access acknowledgements in the gate to open the published PDF preserved in this repository.