Upload 9 files

.gitattributes

@@ -35,3 +35,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 Docker/models/audio_model.keras filter=lfs diff=lfs merge=lfs -text
 Docker/shape_predictor_68_face_landmarks.dat filter=lfs diff=lfs merge=lfs -text
+models/audio_model.keras filter=lfs diff=lfs merge=lfs -text
+shape_predictor_68_face_landmarks.dat filter=lfs diff=lfs merge=lfs -text

Dockerfile

@@ -0,0 +1,36 @@
+# ============================================================
+# GPU-ENABLED SELF-CONTAINED IMAGE FOR DEEPFAKE VIDEO DETECTOR
+# ============================================================
+FROM pytorch/pytorch:2.2.2-cuda12.1-cudnn8-runtime
+
+ENV DEBIAN_FRONTEND=noninteractive \
+    PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1
+
+# Force reinstall Gradio 4.44.1 every time (avoids cache)
+ARG GRADIO_FORCE_REINSTALL=1
+RUN pip install --no-cache-dir --upgrade gradio==4.44.1
+
+
+# System deps for OpenCV + dlib
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential cmake \
+    libgl1 libglib2.0-0 libsm6 libxext6 libxrender1 \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+# Copy app + requirements
+COPY requirements.txt /app/requirements.txt
+COPY app.py /app/app.py
+
+# ✅ Copy your model + landmark file directly into the image
+COPY models /app/models
+COPY shape_predictor_68_face_landmarks.dat /app/
+
+# Install TensorFlow (GPU-capable) + requirements
+RUN pip install --no-cache-dir tensorflow==2.18.0
+RUN pip install --no-cache-dir -r requirements.txt
+
+EXPOSE 7860
+CMD ["python", "app.py"]

app.py

@@ -0,0 +1,237 @@
+import os, io, tempfile, warnings
+import numpy as np
+import gradio as gr
+
+# =========================
+# TensorFlow / Keras
+# =========================
+import tensorflow as tf
+from tensorflow.keras.applications.resnet50 import preprocess_input as resnet_preprocess
+
+# GPU config for TensorFlow (enable if available)
+gpus = tf.config.experimental.list_physical_devices("GPU")
+if gpus:
+    try:
+        for g in gpus:
+            tf.config.experimental.set_memory_growth(g, True)
+        # Mixed precision can speed up on modern GPUs
+        try:
+            tf.keras.mixed_precision.set_global_policy("mixed_float16")
+            print("[INFO] TF mixed precision enabled.")
+        except Exception as e:
+            print(f"[WARN] Could not enable mixed precision: {e}")
+    except Exception as e:
+        print(f"[WARN] Could not set memory growth: {e}")
+
+# =========================
+# Vision stack
+# =========================
+import cv2
+import torch
+from facenet_pytorch import MTCNN
+import dlib
+from imutils import face_utils
+from scipy.spatial import distance as dist
+
+warnings.filterwarnings("ignore")
+
+# =========================
+# Paths / Config
+# =========================
+VIDEO_MODEL_PATH = "models/video_model.h5"
+DLIB_LANDMARK_MODEL = "shape_predictor_68_face_landmarks.dat"
+
+IMG_SIZE = (224, 224)
+FRAME_STEP = 5             # every 5th frame
+NUM_MAX_FACES = 300        # cap on faces collected
+
+# Blink (EAR) features
+EAR_THRESHOLD = 0.25
+EAR_CONSEC_FRAMES = 3
+
+# Prediction threshold
+PRED_THRESHOLD = 0.5
+
+# =========================
+# Lazy-loaded state
+# =========================
+_video_model = None
+_mtcnn = None
+_dlib_detector = None
+_dlib_predictor = None
+
+# Torch / MTCNN device selection
+_torch_device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+if torch.cuda.is_available():
+    torch.backends.cudnn.benchmark = True
+    try:
+        torch.set_float32_matmul_precision("medium")
+    except Exception:
+        pass
+print(f"[INFO] PyTorch device: {_torch_device}")
+
+def lazy_load():
+    global _video_model, _mtcnn, _dlib_detector, _dlib_predictor
+
+    if _video_model is None:
+        if not os.path.exists(VIDEO_MODEL_PATH):
+            raise FileNotFoundError(f"Missing: {VIDEO_MODEL_PATH}")
+        # Allow tf to place on GPU if available
+        _video_model = tf.keras.models.load_model(VIDEO_MODEL_PATH)
+        print("[INFO] Video model loaded.")
+
+    if _mtcnn is None:
+        _mtcnn = MTCNN(keep_all=True, device=_torch_device, image_size=IMG_SIZE[0])
+        print(f"[INFO] MTCNN ready on {_torch_device}.")
+
+    if _dlib_detector is None or _dlib_predictor is None:
+        if not os.path.exists(DLIB_LANDMARK_MODEL):
+            raise FileNotFoundError(
+                f"Missing dlib predictor: {DLIB_LANDMARK_MODEL}. Place it beside app.py."
+            )
+        _dlib_detector = dlib.get_frontal_face_detector()
+        _dlib_predictor = dlib.shape_predictor(DLIB_LANDMARK_MODEL)
+        print("[INFO] dlib detector + predictor ready.")
+
+# =========================
+# VIDEO: faces + blink features
+# =========================
+def _eye_aspect_ratio(eye_pts):
+    A = dist.euclidean(eye_pts[1], eye_pts[5])
+    B = dist.euclidean(eye_pts[2], eye_pts[4])
+    C = dist.euclidean(eye_pts[0], eye_pts[3])
+    return (A + B) / (2.0 * C + 1e-9)
+
+def extract_faces_all(path):
+    """Sample every 5th frame, keep ALL faces per frame, resize to 224x224."""
+    cap = cv2.VideoCapture(path)
+    total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    if total <= 0:
+        cap.release()
+        return None
+
+    faces = []
+    for idx in range(0, total, FRAME_STEP):
+        if len(faces) >= NUM_MAX_FACES:
+            break
+        cap.set(cv2.CAP_PROP_POS_FRAMES, idx)
+        ok, frame = cap.read()
+        if not ok:
+            break
+        rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+        boxes, _ = _mtcnn.detect(rgb)
+        if boxes is None:
+            continue
+        for (x1, y1, x2, y2) in boxes.astype(int):
+            x1, y1 = max(0, x1), max(0, y1)
+            x2, y2 = min(frame.shape[1], x2), min(frame.shape[0], y2)
+            crop = frame[y1:y2, x1:x2]
+            if crop.size > 0:
+                faces.append(cv2.resize(crop, IMG_SIZE))
+            if len(faces) >= NUM_MAX_FACES:
+                break
+    cap.release()
+    if not faces:
+        return None
+    return np.stack(faces, axis=0).astype(np.uint8)
+
+def blink_features_from_crops(faces):
+    """Compute EAR on cropped faces; denominator = #frames with valid EAR."""
+    (lStart, lEnd) = face_utils.FACIAL_LANDMARKS_IDXS["left_eye"]
+    (rStart, rEnd) = face_utils.FACIAL_LANDMARKS_IDXS["right_eye"]
+
+    ear_values = []
+    blink_counter = 0
+    total_blinks = 0
+
+    for face in faces:
+        gray = cv2.cvtColor(face, cv2.COLOR_BGR2GRAY)
+        rects = _dlib_detector(gray, 0)
+        if len(rects) == 0:
+            continue
+        shape = _dlib_predictor(gray, rects[0])
+        shape = face_utils.shape_to_np(shape)
+
+        left_eye = shape[lStart:lEnd]
+        right_eye = shape[rStart:rEnd]
+        ear = 0.5 * (_eye_aspect_ratio(left_eye) + _eye_aspect_ratio(right_eye))
+        ear_values.append(ear)
+
+        if ear < EAR_THRESHOLD:
+            blink_counter += 1
+        else:
+            if blink_counter >= EAR_CONSEC_FRAMES:
+                total_blinks += 1
+            blink_counter = 0
+
+    n_ear = len(ear_values)
+    if n_ear == 0:
+        return np.array([0, 0.0, 0.0], dtype=np.float32)
+
+    blink_freq = total_blinks / float(n_ear)
+    ear_var = float(np.var(np.array(ear_values, dtype=np.float32)))
+    return np.array([total_blinks, blink_freq, ear_var], dtype=np.float32)
+
+def predict_video_prob(video_path):
+    faces = extract_faces_all(video_path)
+    if faces is None or len(faces) == 0:
+        return None, "No faces detected."
+
+    blink_feats = blink_features_from_crops(faces)
+    # Tile to per-face samples
+    tiled = np.tile(blink_feats, (faces.shape[0], 1)).astype(np.float32)
+    imgs = resnet_preprocess(faces.astype(np.float32))
+
+    # Batch size heuristic (bigger if GPU is present)
+    has_tf_gpu = len(tf.config.experimental.list_physical_devices("GPU")) > 0
+    bs = 128 if has_tf_gpu else 32
+
+    preds = []
+    for i in range(0, imgs.shape[0], bs):
+        p = _video_model.predict([imgs[i:i+bs], tiled[i:i+bs]], verbose=0)
+        preds.append(p.reshape(-1))
+    return float(np.mean(np.concatenate(preds, axis=0))), None
+
+def to_verdict(score):
+    return "DEEPFAKE" if score >= PRED_THRESHOLD else "REAL"
+
+# =========================
+# Inference entry
+# =========================
+def run_inference(video_file):
+    lazy_load()
+    if video_file is None:
+        return None, None, "Please upload a video file."
+
+    video_prob, vmsg = predict_video_prob(video_file)
+    if video_prob is None:
+        return None, None, vmsg or "Unable to process the video."
+
+    verdict = f"VIDEO ONLY: {to_verdict(video_prob)}"
+    fmt = lambda x: None if x is None else round(float(x), 4)
+    return fmt(video_prob), verdict, None
+
+# =========================
+# Gradio UI
+# =========================
+with gr.Blocks(title="Deepfake Detector — Video Only (GPU-ready)") as demo:
+    gr.Markdown(
+        "### 🎭 Deepfake Detector — **Video Only**\n"
+        "- **Compute**: Uses GPU automatically if available (PyTorch MTCNN, TensorFlow model; mixed precision on TF).\n"
+        "- **Video path**: samples every 5th frame, keeps **all** faces, resizes to 224×224.\n"
+        "- **Features**: ResNet50 preprocessing + blink EAR features on cropped faces via dlib.\n"
+        f"- **Threshold**: {PRED_THRESHOLD} (≥ means DEEPFAKE).\n"
+    )
+    with gr.Row():
+        video_in = gr.Video(label="Video")
+    go = gr.Button("Analyze")
+    with gr.Row():
+        v_out = gr.Number(label="Video probability (deepfake)", precision=4)
+        verdict_out = gr.Textbox(label="Verdict", interactive=False)
+    msg_out = gr.Textbox(label="Message / Warnings", interactive=False)
+
+    go.click(run_inference, inputs=[video_in], outputs=[v_out, verdict_out, msg_out])
+
+if __name__ == "__main__":
+    lazy_load()
+    demo.launch(server_name="0.0.0.0", server_port=7860)

models/audio_model.keras

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6be7d6e4320dce4e1e74edd0f8f3632aafd7b72be1f0973af72aef7cd52fd08
+size 25015854

models/audio_model/config.json

@@ -0,0 +1 @@
+{"module": "keras", "class_name": "Sequential", "config": {"name": "sequential", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null, "shared_object_id": 139413435787232}, "layers": [{"module": "keras.layers", "class_name": "InputLayer", "config": {"batch_shape": [null, 1000, 55], "dtype": "float32", "sparse": false, "ragged": false, "name": "input_layer"}, "registered_name": null}, {"module": "keras.layers", "class_name": "Conv1D", "config": {"name": "conv1d", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null}, "filters": 64, "kernel_size": [7], "strides": [1], "padding": "valid", "data_format": "channels_last", "dilation_rate": [1], "groups": 1, "activation": "relu", "use_bias": true, "kernel_initializer": {"module": "keras.initializers", "class_name": "GlorotUniform", "config": {"seed": null}, "registered_name": null}, "bias_initializer": {"module": "keras.initializers", "class_name": "Zeros", "config": {}, "registered_name": null}, "kernel_regularizer": null, "bias_regularizer": null, "activity_regularizer": null, "kernel_constraint": null, "bias_constraint": null}, "registered_name": null, "build_config": {"input_shape": [null, 1000, 55]}}, {"module": "keras.layers", "class_name": "MaxPooling1D", "config": {"name": "max_pooling1d", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null, "shared_object_id": 139413435787232}, "pool_size": [2], "padding": "valid", "strides": [2], "data_format": "channels_last"}, "registered_name": null}, {"module": "keras.layers", "class_name": "Dropout", "config": {"name": "dropout", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null, "shared_object_id": 139413435787232}, "rate": 0.30000000000000004, "seed": null, "noise_shape": null}, "registered_name": null}, {"module": "keras.layers", "class_name": "Conv1D", "config": {"name": "conv1d_1", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null, "shared_object_id": 139413435787232}, "filters": 128, "kernel_size": [5], "strides": [1], "padding": "valid", "data_format": "channels_last", "dilation_rate": [1], "groups": 1, "activation": "relu", "use_bias": true, "kernel_initializer": {"module": "keras.initializers", "class_name": "GlorotUniform", "config": {"seed": null}, "registered_name": null}, "bias_initializer": {"module": "keras.initializers", "class_name": "Zeros", "config": {}, "registered_name": null}, "kernel_regularizer": null, "bias_regularizer": null, "activity_regularizer": null, "kernel_constraint": null, "bias_constraint": null}, "registered_name": null, "build_config": {"input_shape": [null, 497, 64]}}, {"module": "keras.layers", "class_name": "MaxPooling1D", "config": {"name": "max_pooling1d_1", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null, "shared_object_id": 139413435787232}, "pool_size": [2], "padding": "valid", "strides": [2], "data_format": "channels_last"}, "registered_name": null}, {"module": "keras.layers", "class_name": "Dropout", "config": {"name": "dropout_1", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null, "shared_object_id": 139413435787232}, "rate": 0.30000000000000004, "seed": null, "noise_shape": null}, "registered_name": null}, {"module": "keras.layers", "class_name": "Flatten", "config": {"name": "flatten", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null, "shared_object_id": 139413435787232}, "data_format": "channels_last"}, "registered_name": null, "build_config": {"input_shape": [null, 246, 128]}}, {"module": "keras.layers", "class_name": "Dense", "config": {"name": "dense", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null, "shared_object_id": 139413435787232}, "units": 64, "activation": "relu", "use_bias": true, "kernel_initializer": {"module": "keras.initializers", "class_name": "GlorotUniform", "config": {"seed": null}, "registered_name": null}, "bias_initializer": {"module": "keras.initializers", "class_name": "Zeros", "config": {}, "registered_name": null}, "kernel_regularizer": null, "bias_regularizer": null, "kernel_constraint": null, "bias_constraint": null}, "registered_name": null, "build_config": {"input_shape": [null, 31488]}}, {"module": "keras.layers", "class_name": "Dense", "config": {"name": "dense_1", "trainable": true, "dtype": {"module": "keras", "class_name": "DTypePolicy", "config": {"name": "float32"}, "registered_name": null, "shared_object_id": 139413435787232}, "units": 1, "activation": "sigmoid", "use_bias": true, "kernel_initializer": {"module": "keras.initializers", "class_name": "GlorotUniform", "config": {"seed": null}, "registered_name": null}, "bias_initializer": {"module": "keras.initializers", "class_name": "Zeros", "config": {}, "registered_name": null}, "kernel_regularizer": null, "bias_regularizer": null, "kernel_constraint": null, "bias_constraint": null}, "registered_name": null, "build_config": {"input_shape": [null, 64]}}], "build_input_shape": [null, 1000, 55]}, "registered_name": null, "build_config": {"input_shape": [null, 1000, 55]}, "compile_config": {"optimizer": {"module": "keras.optimizers", "class_name": "Adam", "config": {"name": "adam", "learning_rate": 0.0010000000474974513, "weight_decay": null, "clipnorm": null, "global_clipnorm": null, "clipvalue": null, "use_ema": false, "ema_momentum": 0.99, "ema_overwrite_frequency": null, "loss_scale_factor": null, "gradient_accumulation_steps": null, "beta_1": 0.9, "beta_2": 0.999, "epsilon": 1e-07, "amsgrad": false}, "registered_name": null}, "loss": "binary_crossentropy", "loss_weights": null, "metrics": ["accuracy"], "weighted_metrics": null, "run_eagerly": false, "steps_per_execution": 1, "jit_compile": false}}

models/audio_model/metadata.json

@@ -0,0 +1 @@
+{"keras_version": "3.10.0", "date_saved": "2025-09-30@02:28:41"}

models/audio_model/model.weights.h5

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e9a4c54ac044bcb96c43b9947c05af09fc1de6cceb3f2881a5761a2b2da8768c
+size 25009192

models/video_model.h5

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f623abf62c659bcfd8364bcc4b233680b6834d5371b925bfe62a9859efb7bd39
+size 285090744

requirements.txt

@@ -0,0 +1,30 @@
+# -----------------------------
+# Core scientific stack
+# -----------------------------
+numpy==1.26.4
+scipy==1.11.4
+pandas==2.2.2
+
+# -----------------------------
+# TensorFlow (GPU-enabled if CUDA present)
+# -----------------------------
+tensorflow==2.18.0
+
+# -----------------------------
+# PyTorch + facenet-pytorch
+# (CPU wheels by default; CUDA wheels provided in Dockerfile)
+# -----------------------------
+facenet-pytorch==2.6.0
+pillow==10.2.0
+
+# -----------------------------
+# Vision / utility
+# -----------------------------
+opencv-python-headless==4.9.0.80
+imutils==0.5.4
+dlib==19.24.4
+
+# -----------------------------
+# Web UI
+# -----------------------------
+gradio==4.4.1

shape_predictor_68_face_landmarks.dat

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fbdc2cb80eb9aa7a758672cbfdda32ba6300efe9b6e6c7a299ff7e736b11b92f
+size 99693937