Dwiyuda/KTM-Detector-UIR-Skripsi_Public

Tugas Akhir Teknik Informatika UIR 2026, Sistem deteksi otomatis keaslian Kartu Tanda Mahasiswa (KTM) berbasis web. Pipeline: Qwen3-VL (gerbang semantik) → Florence-2 + Canny (deteksi & crop kartu) → EfficientNet-B0 (klasifikasi ASLI/PALSU) → Grad-CAM. Dataset 1.645 citra dengan 5 skenario pemalsuan terkontrol. # KTM Detector UIR — Automatic Student ID Card Authenticity Detection Web-based system that determines whether a **Kartu Tanda Mahasiswa** (student ID card) of Universitas Islam Riau is **genuine** or **forged**, using a Convolutional Neural Network. Undergraduate thesis — Informatics Engineering, Universitas Islam Riau, 2026 **Dwi Yuda** · NPM 223510389 *Baca dalam Bahasa Indonesia →* --- ## How It Works ``` Uploaded photo │ ├─ 1. Qwen3-VL gate "Is there an ID card in this image?" │ answer "NO" → REJECT (classifier never runs) │ ├─ 2. Florence-2 + Canny locate, crop and de-skew the card → 1024×474 px │ ├─ 3. EfficientNet-B0 classify REAL / FAKE │ └─ 4. Grad-CAM highlight the regions behind the decision ``` The system is **self-contained**: it needs no reference database of genuine cards. The model learns pixel-level forensic traces — damaged guilloché patterns, typographic inconsistency, print-and-rescan artefacts — that are invisible to the naked eye. --- ## Results ### Dataset 1,645 images: **845 genuine** + **800 forged** across five controlled attack scenarios. | Scenario | Count | Forgery type | |----------|-------|--------------| | A | 200 | Face photo replaced by generative AI (Stable Diffusion) | | B | 200 | Typographic inconsistency in identity text | | C | 80 | Print-and-rescan recapture | | D | 200 | Identity text forged (EasyOCR + copy-patch, guilloché preserved) | | Hybrid | 120 | Combination of A/B/D followed by print-and-rescan | Every forged image derives from a genuine card with the same number (`real_N` → `fake_N`), so the train/validation/test split is **group-aware**: all derivatives of one physical card stay in the same subset. This prevents card-level data leakage. ### Classification performance At the architectures' **native input resolution** (224×224 for B0, 300×300 for B3): | Model | Split | Accuracy | Precision | Recall | F1 | ROC-AUC | |-------|-------|----------|-----------|--------|--- …