ARGUS: Identity Document Fraud Detection

A production-grade AI system designed to detect next-generation identity document fraud, built for the FREUID Challenge 2026 (hosted at IJCAI-ECAI 2026).

ARGUS is designed to address three primary types of document forgery:

1. Physical Manipulations: Alterations made directly on physical document substrates.
2. GenAI-Driven Edits: Digital manipulations generated using generative AI tools.
3. Print-and-Capture (Recapture) Attacks: Forged documents that are printed and re-photographed to mask digital editing artifacts.

---

Project Status

This repository is the reproducibility package for Team ARGUS's FREUID Challenge 2026 submission. The system architecture, compliance checkpoints, and delivery history are documented in the Enterprise AI Project Runbook.

The Docker image ships with a trained EfficientNet-B4 champion checkpoint (checkpoints/champion_efficientnet_b4.pth, produced by scripts/train_champion_checkpoint.sh) baked in — see Reproducing the Submission below to rebuild and verify it end-to-end.

Repository Structure

ARGUS/
├── docs/                           # Architecture and design documentation
│   ├── project_runbook.md          # End-to-end project runbook
│   ├── 10_Product_Backlog.md       # Product backlog
│   ├── Phase-0/                    # Charter, stakeholder matrix, architecture vision
│   │   ├── 00_Project_Charter.md
│   │   └── 01_Architecture_Vision.md
│   ├── Phase-1/                    # Business & use-case requirements
│   │   ├── 02_BRD.md
│   │   └── 03_Use_Case_Specification.md
│   ├── Phase-2/                    # Solution & data architecture
│   │   ├── 04_SAD.md
│   │   ├── 05_DAD.md
│   │   └── adr/                    # Architecture Decision Records
│   ├── Phase-3/                    # AI/ML design & test strategy
│   │   ├── 06_ML_Design.md
│   │   └── 07_Test_Strategy.md
│   └── Phase-4/                    # Security, compliance & operations
│       ├── 08_Security_Compliance.md
│       └── 09_Operations_Runbook.md
├── src/                            # Core source code
│   ├── data/                       # Preprocessing and augmentation pipelines
│   ├── models/                     # Model definitions and ensembles
│   └── training/                   # Training and evaluation scripts
├── configs/                        # Hydra configuration files
└── notebooks/                      # Exploratory data analysis (EDA)

---

Technical Approach

Model Architecture

ARGUS is built around an ensemble of diverse deep learning backbones that trade off latency and accuracy:

• EVA-02-Large / ConvNeXt-V2-Base: High-capacity feature extractors for fine-grained texture and pattern analysis.
• EfficientNet-B4: A lightweight backbone optimized for lower-latency inference — this is the backbone shipped as the Docker image's default champion checkpoint, since it is the fastest to train and serve while the full three-backbone ensemble (--model_name ensemble) remains available for teams with the compute budget to train all three.

Evaluation Metrics

We optimize for the competition's primary metric along with key production-focused metrics:

• AuDET (Area under the Detection Error Trade-off curve): Evaluates overall system trade-offs.
• APCER @ 1% BPCER: Measures the rate of fraudulent documents accepted when the false rejection rate of genuine documents is capped at 1%.

---

Quick Start

1. Setup Environment

git clone https://github.com/mittalpk/ARGUS
cd ARGUS
bash scripts/setup.sh

2. Dataset Acquisition

Ensure your Kaggle API credentials are configured in ~/.kaggle/kaggle.json or a local .env (see .env — never commit real credentials):

bash scripts/download_data.sh

---

Author

Praveen Mittal LinkedIn

---

License

Licensed under the Apache 2.0 License. See LICENSE for details.