Initial release: Unified Dual-Expert ViT balanced AI detector

README.md

@@ -0,0 +1,27 @@
+---
+language:
+- en
+- id
+license: apache-2.0
+tags:
+- vision
+- image-classification
+- deepfake-detection
+- ai-generated
+- defactify
+- dual-expert
+pipeline_tag: image-classification
+---
+
+# Dual-Expert Vision Transformer (Defactify Balanced)
+
+Arsitektur ensemble spesialis untuk memecahkan classifier bias pada deteksi gambar AI modern.
+- **Pakar Real (Guardian):** Mencegah false positive pada foto asli (Weight [3.0, 1.0]).
+- **Pakar AI (Hunter):** Menangkap artefak generator difusi modern (Weight [1.0, 3.0]).
+
+## Benchmark Evaluasi (500 Sampel Defactify Test)
+- Balanced Accuracy: 79.00%
+- Macro F1: 79.00%
+- Recall Real: 80.00%
+- Recall AI: 78.00%
+- Latency: ~6.57 ms (Concurrent CUDA Streams)

config.json

@@ -0,0 +1,27 @@
+{
+  "model_type": "dual_expert_vision_transformer",
+  "base_architecture": "google/vit-base-patch16-224",
+  "num_classes": 2,
+  "id2label": {
+    "0": "Real Camera",
+    "1": "AI-Generated"
+  },
+  "label2id": {
+    "Real Camera": 0,
+    "AI-Generated": 1
+  },
+  "calibration": {
+    "method": "Linear Weighted Consensus",
+    "w_hunter": 0.6,
+    "w_guardian": 0.4,
+    "threshold": 0.46
+  },
+  "metrics_benchmark": {
+    "dataset": "Defactify Balanced (500 Test)",
+    "accuracy": 0.79,
+    "macro_f1": 0.79,
+    "recall_real": 0.8,
+    "recall_ai": 0.78,
+    "latency_ms": 6.57
+  }
+}

guardian/config.json

@@ -0,0 +1,24 @@
+{
+  "architectures": [
+    "ViTForImageClassification"
+  ],
+  "attention_probs_dropout_prob": 0.0,
+  "dtype": "float32",
+  "encoder_stride": 16,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.0,
+  "hidden_size": 768,
+  "image_size": 224,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-12,
+  "model_type": "vit",
+  "num_attention_heads": 12,
+  "num_channels": 3,
+  "num_hidden_layers": 12,
+  "patch_size": 16,
+  "pooler_act": "tanh",
+  "pooler_output_size": 768,
+  "qkv_bias": true,
+  "transformers_version": "5.0.0"
+}

guardian/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:697950f2f90f29120db62f55dbfce8ef801b28adff7ea6e05b5528886b5a5945
+size 343223968

hunter/config.json

@@ -0,0 +1,24 @@
+{
+  "architectures": [
+    "ViTForImageClassification"
+  ],
+  "attention_probs_dropout_prob": 0.0,
+  "dtype": "float32",
+  "encoder_stride": 16,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.0,
+  "hidden_size": 768,
+  "image_size": 224,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-12,
+  "model_type": "vit",
+  "num_attention_heads": 12,
+  "num_channels": 3,
+  "num_hidden_layers": 12,
+  "patch_size": 16,
+  "pooler_act": "tanh",
+  "pooler_output_size": 768,
+  "qkv_bias": true,
+  "transformers_version": "5.0.0"
+}

hunter/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e25d011f8ce70594c1a6577fdbbb735592c671008293e816d5e909b86dcbdba4
+size 343223968

pipeline.py

@@ -0,0 +1,35 @@
+import os
+import json
+import torch
+import torch.nn as nn
+from transformers import AutoImageProcessor, AutoModelForImageClassification
+from PIL import Image
+
+class DualExpertDetector(nn.Module):
+    def __init__(self, model_dir: str):
+        super().__init__()
+        self.processor = AutoImageProcessor.from_pretrained(model_dir)
+        self.guardian = AutoModelForImageClassification.from_pretrained(os.path.join(model_dir, 'guardian'))
+        self.hunter = AutoModelForImageClassification.from_pretrained(os.path.join(model_dir, 'hunter'))
+        with open(os.path.join(model_dir, 'config.json'), 'r') as f:
+            cfg = json.load(f)['calibration']
+        self.w_hunter = cfg['w_hunter']
+        self.w_guardian = cfg['w_guardian']
+        self.threshold = cfg['threshold']
+
+    def predict(self, image: Image.Image, device: str = 'cuda'):
+        self.to(device)
+        self.eval()
+        inputs = self.processor(images=image.convert('RGB'), return_tensors='pt').to(device)
+        with torch.no_grad():
+            p_g = torch.softmax(self.guardian(**inputs).logits, dim=-1)[0, 1].item()
+            p_h = torch.softmax(self.hunter(**inputs).logits, dim=-1)[0, 1].item()
+        combined = (self.w_hunter * p_h) + (self.w_guardian * p_g)
+        is_ai = combined >= self.threshold
+        return {
+            'verdict': 'AI-Generated' if is_ai else 'Real Camera',
+            'is_ai': bool(is_ai),
+            'confidence_ai': round(combined * 100, 2),
+            'confidence_real': round((1.0 - combined) * 100, 2),
+            'raw_scores': {'hunter_ai': round(p_h * 100, 2), 'guardian_ai': round(p_g * 100, 2)}
+        }

preprocessor_config.json

@@ -0,0 +1,23 @@
+{
+  "do_convert_rgb": null,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "image_processor_type": "ViTImageProcessor",
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "resample": 2,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "height": 224,
+    "width": 224
+  }
+}