🛡️ SecurePixel: Hunter Model (Trained from Scratch)
Model forensik Hunter berbasis arsitektur ConvNeXtV2-Tiny yang dilatih murni dari awal (from scratch / zero pretraining) untuk mendeteksi citra sintetis buatan AI vs citra sensor kamera riil.
⚙️ Spesifikasi & Protokol Training:
- Arsitektur:
convnextv2_tiny(Global Response Normalization) - Tipe Pelatihan: From Scratch (Bobot murni diinisialisasi acak)
- Presisi: FP32 Murni (Full Precision, tanpa kuantisasi)
- Input Resolution: Pure 512x512 non-overlapping crop (No resizing / No interpolation)
- Augmentasi Dinamis: On-the-fly real-world compression (JPEG & WebP social media degradation)
- Dataset Parity: 1:1 Strict Parity per batch:
- Class 1 (AI): Kisi laten generator modern (Defactify, Imagen/Gemini, Midjourney, SDXL, FLUX).
- Class 0 (Real): Sensor fisik optik CMOS kamera nyata (Sensor D05–D40).