""" Vision Transformer (ViT) Deepfake Detector Adapter Provides a clean, standardized wrapper around the dima806/deepfake_vs_real_image_detection ViT model. Adheres strictly to the DetectorResult schema, supporting lazy loading and robust error handling. """ import os import sys import time from typing import Optional, Union, Any, Dict from PIL import Image import torch from transformers import AutoImageProcessor, AutoModelForImageClassification from src.detector_schema import ( DetectorResult, PREDICTION_REAL, PREDICTION_FAKE, PREDICTION_ERROR, ) BASE_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) LOCAL_MODEL_DIR = os.path.join(BASE_DIR, "models", "deepfake_vs_real_image_detection") HUB_MODEL_ID = "dima806/deepfake_vs_real_image_detection" class ViTAdapter: """ Adapter for dima806 ViT Deepfake Detection Model. Model Details: - Architecture: ViTForImageClassification (google/vit-base-patch16-224 fine-tuned) - Checkpoint: models/deepfake_vs_real_image_detection (model.safetensors) - Image Preprocessing: ViTImageProcessor (Resize 224x224, Rescale 1/255, Normalize mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) - Decision Threshold: 0.50 Fake Probability - Label Mapping: 0 = Real, 1 = Fake """ DETECTOR_NAME = "ViTDeepfakeDetector" MODEL_NAME = "dima806/deepfake_vs_real_image_detection (ViT-Base/16)" PREPROCESSING_NAME = "ViTImageProcessor(Resize 224x224, Rescale 1/255, Normalize mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])" DECISION_THRESHOLD = 0.50 LIMITATION_WARNING = ( "Vision Transformer trained for facial deepfake detection. " "Evaluates learned visual artifact representations; does not guarantee authenticity." ) def __init__(self, device: Optional[Union[str, torch.device]] = None, lazy_load: bool = True): self.device_requested = device self.processor = None self.model = None self.device = None self.real_idx = 0 self.fake_idx = 1 self.load_error = None self.load_time_sec = None if not lazy_load: self.load_model() def load_model(self) -> bool: """Loads the ViT processor and classifier weights once if not already initialized.""" if self.model is not None: return True start_t = time.perf_counter() try: # Determine compute device if self.device_requested is not None: self.device = torch.device(self.device_requested) else: self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # Determine model path (local directory preferred) if os.path.exists(LOCAL_MODEL_DIR) and os.path.exists(os.path.join(LOCAL_MODEL_DIR, "config.json")): model_source = LOCAL_MODEL_DIR local_only = True else: model_source = HUB_MODEL_ID local_only = False self.processor = AutoImageProcessor.from_pretrained( model_source, local_files_only=local_only ) self.model = AutoModelForImageClassification.from_pretrained( model_source, local_files_only=local_only ) self.model.to(self.device) self.model.eval() # Verify label indices dynamically from model config config_id2label = {int(k): v.lower() for k, v in self.model.config.id2label.items()} self.real_idx = next(k for k, v in config_id2label.items() if "real" in v) self.fake_idx = next(k for k, v in config_id2label.items() if "fake" in v) self.load_time_sec = time.perf_counter() - start_t return True except Exception as e: self.load_error = str(e) self.processor = None self.model = None return False def unload_model(self) -> None: """ Explicitly releases the Vision Transformer model and image processor to free system and GPU memory. """ self.model = None self.processor = None if torch.cuda.is_available(): torch.cuda.empty_cache() def predict(self, image_input: Any) -> DetectorResult: """ Execute single-image inference and return a standardized DetectorResult. Accepts: - PIL Image - File path string - Numpy image array (or any object convertible to PIL Image) """ start_time = time.perf_counter() # 1. Validation for empty input if image_input is None: return DetectorResult.create_error( detector_name=self.DETECTOR_NAME, model_name=self.MODEL_NAME, error_message="No image provided (input was None).", warning=self.LIMITATION_WARNING, device=str(self.device) if self.device else None, preprocessing=self.PREPROCESSING_NAME, ) # 2. Safe image loading and RGB conversion try: if isinstance(image_input, str): if not os.path.exists(image_input): return DetectorResult.create_error( detector_name=self.DETECTOR_NAME, model_name=self.MODEL_NAME, error_message=f"Image file does not exist: {image_input}", warning=self.LIMITATION_WARNING, device=str(self.device) if self.device else None, preprocessing=self.PREPROCESSING_NAME, ) with Image.open(image_input) as opened_img: img_rgb = opened_img.convert("RGB") elif isinstance(image_input, Image.Image): img_rgb = image_input.convert("RGB") elif hasattr(image_input, "__array__"): img_rgb = Image.fromarray(image_input).convert("RGB") else: return DetectorResult.create_error( detector_name=self.DETECTOR_NAME, model_name=self.MODEL_NAME, error_message=f"Unsupported image input type: {type(image_input).__name__}", warning=self.LIMITATION_WARNING, device=str(self.device) if self.device else None, preprocessing=self.PREPROCESSING_NAME, ) except Exception as img_err: return DetectorResult.create_error( detector_name=self.DETECTOR_NAME, model_name=self.MODEL_NAME, error_message=f"Invalid or corrupted image: {str(img_err)}", warning=self.LIMITATION_WARNING, device=str(self.device) if self.device else None, preprocessing=self.PREPROCESSING_NAME, ) # 3. Ensure model is loaded if not self.load_model(): return DetectorResult.create_error( detector_name=self.DETECTOR_NAME, model_name=self.MODEL_NAME, error_message=f"Model loading failed: {self.load_error}", warning=self.LIMITATION_WARNING, device=str(self.device) if self.device else None, preprocessing=self.PREPROCESSING_NAME, ) # 4. Inference execution under torch.inference_mode() try: # Dynamic device adjustment if running inside ZeroGPU context current_device = torch.device("cuda" if torch.cuda.is_available() else "cpu") if next(self.model.parameters()).device != current_device: self.model.to(current_device) self.device = current_device inputs = self.processor(images=img_rgb, return_tensors="pt") inputs = {k: v.to(self.device) for k, v in inputs.items()} with torch.inference_mode(): outputs = self.model(**inputs) logits = outputs.logits # shape [1, 2] probs = torch.softmax(logits, dim=1).squeeze() fake_prob = float(probs[self.fake_idx].item()) real_prob = float(probs[self.real_idx].item()) # Relative logit (fake_logit - real_logit), such that logit >= 0 aligns with prob >= 0.50 real_logit = float(logits[0, self.real_idx].item()) fake_logit = float(logits[0, self.fake_idx].item()) relative_logit = fake_logit - real_logit latency_ms = (time.perf_counter() - start_time) * 1000 # Output decision and confidence is_fake = fake_prob >= self.DECISION_THRESHOLD prediction = PREDICTION_FAKE if is_fake else PREDICTION_REAL confidence = fake_prob if is_fake else real_prob return DetectorResult( detector_name=self.DETECTOR_NAME, model_name=self.MODEL_NAME, prediction=prediction, fake_probability=round(fake_prob, 6), real_probability=round(real_prob, 6), raw_logit=round(relative_logit, 6), confidence=round(confidence, 6), threshold=self.DECISION_THRESHOLD, device=str(self.device), preprocessing=self.PREPROCESSING_NAME, inference_time_ms=round(latency_ms, 2), error=None, warning=self.LIMITATION_WARNING, ) except Exception as inf_err: return DetectorResult.create_error( detector_name=self.DETECTOR_NAME, model_name=self.MODEL_NAME, error_message=f"Inference execution failed: {str(inf_err)}", warning=self.LIMITATION_WARNING, device=str(self.device) if self.device else None, preprocessing=self.PREPROCESSING_NAME, )