File size: 3,530 Bytes
f64d8fb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import json
import base64
import numpy as np
import audioop
from scipy.signal import resample
import onnxruntime as ort
from fastapi import FastAPI, WebSocket
import gradio as gr

app = FastAPI()

session = ort.InferenceSession("model.onnx")

@app.get("/")
def home():
    return {"status": "Fraud Shield AI Engine is Active via Gradio Wrapper"}

# 3. Real-Time Telephony WebSocket Endpoint
@app.websocket("/media-stream")
async def handle_twilio_stream(websocket: WebSocket):
    await websocket.accept()
    print("🚀 Twilio Call Connected via Live Media Stream WebSocket")
    
    audio_buffer = bytearray()
    
    try:
        while True:
            # Twilio streams packets continuously over the WebSocket
            message = await websocket.receive_text()
            data = json.loads(message)
            
            if data['event'] == 'media':
                # Extract the 20ms base64 chunk
                payload = data['media']['payload']
                chunk = base64.b64decode(payload)
                
                # Step A: Convert 8-bit Mu-Law telecom audio to 16-bit Linear PCM
                pcm_data = audioop.ulaw2lin(chunk, 2)
                audio_buffer.extend(pcm_data)
                
                # Process audio continuously in 1-second sliding windows (16,000 bytes at 8kHz)
                if len(audio_buffer) >= 16000:
                    raw_audio = np.frombuffer(audio_buffer, dtype=np.int16).astype(np.float32)
                    audio_buffer = bytearray() # Reset buffer for the next sliding frame
                    
                    # Step B: Standardize the magnitude bounds
                    max_val = np.max(np.abs(raw_audio))
                    if max_val > 0:
                        raw_audio /= max_val
                    
                    # Step C: Upsample the 8kHz Twilio track to the model's native 16kHz structure
                    num_samples = int(len(raw_audio) * (16000 / 8000))
                    resampled_audio = resample(raw_audio, num_samples).astype(np.float32)
                    
                    # Step D: Run ONNX Real-Time Inference
                    inputs = {session.get_inputs()[0].name: np.expand_dims(resampled_audio, axis=0)}
                    outputs = session.run(None, inputs)
                    logits = outputs[0]
                    
                    # Step E: Apply Softmax to extract the absolute Deepfake Confidence Score
                    exp_logits = np.exp(logits - np.max(logits))
                    probabilities = exp_logits / np.sum(exp_logits)
                    
                    # Index 0 = Real (Bona Fide), Index 1 = Fake (Synthetic Deepfake)
                    fake_confidence = float(probabilities[0][1])
                    
                    # Broadcast the real-time score out to the terminal console
                    verdict = "⚠️ DEEPFAKE DETECTED" if fake_confidence > 0.80 else "✅ HUMAN VERIFIED"
                    print(f"[{verdict}] Confidence Score: {fake_confidence * 100:.2f}%")
                    
    except Exception as e:
        print(f"Call disconnected or stream ended: {e}")
    finally:
        await websocket.close()


with gr.Blocks() as ui:
    gr.Markdown("# 🛡️ Fraud Shield Live Deepfake Detection Backend")
    gr.Markdown("The background server is up and listening for Twilio WebSocket pipelines.")

app = gr.routes.mount_gradio_app(app, ui, path="/")