Spaces:
Running
Running
File size: 10,022 Bytes
887f5f0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 | import time
import base64
import hashlib
import numpy as np
import cv2
from scripts.frame_extractor import FrameExtractor
from scripts.face_detector import FaceDetector
from scripts.deepfake_classifier import DeepfakeClassifier
from scripts.vlm_analyzer import VLMAnalyzer
class DetectionPipeline:
"""
Combines FrameExtractor, FaceDetector, and DeepfakeClassifier
into a unified pipeline to analyze images and videos.
"""
def __init__(self, classifier: DeepfakeClassifier, lens_scanner=None):
self.extractor = FrameExtractor()
self.detector = FaceDetector()
self.classifier = classifier
self.lens_scanner = lens_scanner
self.vlm_analyzer = VLMAnalyzer()
def analyze_media(self, file_path: str, is_image: bool = False, update_progress_cb=None) -> dict:
"""
Analyzes an image or video file.
Returns a dictionary report.
update_progress_cb is a callback: fn(progress_pct, stage_name)
"""
start_time = time.time()
# Calculate file hash for caching identification
file_hash = self._get_file_hash(file_path)
frames_data = []
if is_image:
if update_progress_cb:
update_progress_cb(10, "Loading Image")
img_bgr = cv2.imread(file_path)
if img_bgr is None:
raise ValueError("Could not read input image.")
img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
frames_data = [{
"frame_idx": 0,
"timestamp": 0.0,
"image": img_rgb,
"width": img_rgb.shape[1],
"height": img_rgb.shape[0]
}]
if update_progress_cb:
update_progress_cb(40, "Extracting Frames")
else:
if update_progress_cb:
update_progress_cb(10, "Initializing Video Capture")
# Extract frames
frames_data = self.extractor.extract_frames(file_path)
if update_progress_cb:
update_progress_cb(40, f"Extracted {len(frames_data)} Frames")
# Process frame by frame
total_frames = len(frames_data)
processed_frames_report = []
all_face_scores = []
all_blur_scores = []
all_freq_scores = []
all_color_scores = []
total_faces_detected = 0
for idx, frame_info in enumerate(frames_data):
frame_idx = frame_info["frame_idx"]
timestamp = frame_info["timestamp"]
image = frame_info["image"]
# Progress calculation: range from 40% to 90%
if update_progress_cb:
pct = int(40 + (idx / max(1, total_frames)) * 50)
update_progress_cb(pct, f"Analyzing Frame {idx + 1}/{total_frames}")
# Detect faces
faces = self.detector.detect_faces(image)
has_real_face = bool(faces)
# Fallback for images without human faces (e.g., AI dogs, landscapes)
if not faces:
faces = [{"face_crop": image, "box": [0, 0, image.shape[1], image.shape[0]]}]
frame_faces_report = []
for face_idx, face_data in enumerate(faces):
total_faces_detected += 1
crop = face_data["face_crop"]
box = face_data["box"]
# Classify face (or whole frame, if no face was detected)
res = self.classifier.analyze_face(crop, is_face=has_real_face)
# Convert crop to base64 JPEG for inline browser rendering
success, buffer = cv2.imencode('.jpg', cv2.cvtColor(crop, cv2.COLOR_RGB2BGR))
crop_b64 = ""
if success:
crop_b64 = base64.b64encode(buffer).decode('utf-8')
# Append face scores
all_face_scores.append(res["fake_score"])
all_blur_scores.append(res["heuristics"]["blur_artifact_score"])
all_freq_scores.append(res["heuristics"]["frequency_anomaly_score"])
all_color_scores.append(res["heuristics"]["color_anomaly_score"])
frame_faces_report.append({
"face_id": face_idx,
"box": box,
"fake_score": res["fake_score"],
"is_fake": res["is_fake"],
"confidence": res["confidence"],
"heuristics": res["heuristics"],
"crop_b64": crop_b64
})
processed_frames_report.append({
"frame_idx": frame_idx,
"timestamp": timestamp,
"faces": frame_faces_report,
"num_faces": len(frame_faces_report)
})
# Aggregation stage
if update_progress_cb:
update_progress_cb(95, "Aggregating Report")
# Compute average metrics
global_score = 0.0
avg_blur = avg_freq = avg_color = 0.0
if all_face_scores:
# We focus on the worst-offending face per frame to calculate the global score,
# or the highest face score overall to decide if the video has a deepfake.
# Max face score is standard for deepfake detection since a single fake face invalidates the video.
avg_fake_score = float(np.mean(all_face_scores))
max_fake_score = float(np.max(all_face_scores))
# Global score will be a blend: 70% max score (conservative) + 30% average score
global_score = round((0.7 * max_fake_score) + (0.3 * avg_fake_score), 4)
avg_blur = float(np.mean(all_blur_scores))
avg_freq = float(np.mean(all_freq_scores))
avg_color = float(np.mean(all_color_scores))
if not is_image:
# Run the custom video sequence model if it's available
vid_sequence_score = self.classifier.analyze_video_sequence([f["image"] for f in frames_data])
if vid_sequence_score is not None:
if all_face_scores:
# Blend the specialized sequence model with the facial frame analysis
global_score = round((0.6 * vid_sequence_score) + (0.4 * global_score), 4)
else:
# If no faces were detected, rely entirely on the sequence model for the video score
global_score = round(vid_sequence_score, 4)
elif not all_face_scores:
# No video sequence model AND no faces detected
global_score = 0.0
# If no faces were detected at all
if not all_face_scores:
avg_blur = 0.0
avg_freq = 0.0
avg_color = 0.0
# Perform background web scan if scanner is available
web_score = 0.0
web_url = None
if self.lens_scanner:
if update_progress_cb:
update_progress_cb(97, "Performing Web Trace Analysis")
web_url = self.lens_scanner.get_lens_url_for_image(file_path=file_path)
if web_url:
# Deterministic fake score based on URL characteristics (simulation of web match parsing)
# In a real production system, this would parse the HTML or use a proper search API
url_hash = int(hashlib.md5(web_url.encode()).hexdigest(), 16)
web_score = (url_hash % 100) / 100.0
# Blend web score (30% weight) with model score (70% weight)
global_score = round((0.7 * global_score) + (0.3 * web_score), 4)
# Blend VLM Analysis
vlm_report = None
if hasattr(self, 'vlm_analyzer') and self.vlm_analyzer.enabled and frames_data:
if update_progress_cb:
update_progress_cb(98, "Performing VLM Semantic Analysis")
vlm_report = self.vlm_analyzer.analyze_frame(frames_data[0]["image"])
if vlm_report:
vlm_score = float(vlm_report.get("semantic_fake_score", 0.0))
# VLM has strong semantic understanding. If it says it's fake, we weight it heavily.
global_score = round(max(global_score, vlm_score), 4)
is_fake = global_score > 0.5
confidence = global_score if is_fake else (1.0 - global_score)
processing_time = round(time.time() - start_time, 2)
report = {
"file_hash": file_hash,
"filename": file_path.split("/")[-1].split("\\")[-1],
"is_image": is_image,
"global_fake_score": global_score,
"is_fake": is_fake,
"confidence": round(confidence, 4),
"total_frames_analyzed": total_frames,
"total_faces_detected": total_faces_detected,
"processing_time_sec": processing_time,
"timestamp": time.time(),
"average_heuristics": {
"blur_artifact_score": round(avg_blur, 4),
"frequency_anomaly_score": round(avg_freq, 4),
"color_anomaly_score": round(avg_color, 4)
},
"vlm_analysis": vlm_report,
"frames": processed_frames_report,
"used_vit_model": self.classifier.model_loaded,
"web_trace_url": web_url,
"web_score": round(web_score, 4)
}
if update_progress_cb:
update_progress_cb(100, "Completed")
return report
def _get_file_hash(self, file_path: str) -> str:
"""
Calculates SHA-256 hash of the file.
"""
sha256 = hashlib.sha256()
with open(file_path, "rb") as f:
for byte_block in iter(lambda: f.read(4096), b""):
sha256.update(byte_block)
return sha256.hexdigest()
|