How To Do AI Real-Time Deepfakes

Real-Time Deepfake Guide

Real-time deepfake is the most dangerous vector — live video calls where the target interacts with a fake version of someone they trust. Here's a(n) article / guide on how to do AI real-time deepfakes.

The Complete Real-Time Pipeline

The full real-time pipeline has three components running simultaneously:

[Your Webcam] → [Face Detection + Swap Engine] → [Virtual Camera Output]
[Your Mic]    → [Voice Cloner + Pitch Shifter]  → [Virtual Audio Output]

All delivered to the target via Zoom / Teams / Meet / Slack / FaceTime.

Hardware Requirements

Component	Minimum	Recommended
GPU	NVIDIA GTX 1060 6GB	RTX 3070+ / RTX 4090
RAM	16GB	32GB
CPU	Recent i5/Ryzen 5	i7/Ryzen 7
Webcam	1080p	4K (allows downscaling)
Monitor	Single	Dual (one for control panel)

Real-Time Face Swap: Deep-Live-Cam

This is currently the best open-source tool for real-time deepfake.

Installation

bash

git clone https://github.com/hacksider/Deep-Live-Cam.git
cd Deep-Live-Cam

# Install all dependencies
python install.py

# If you need CUDA:
pip uninstall onnxruntime onnxruntime-gpu
pip install onnxruntime-gpu --upgrade

Usage

bash

python run.py --execution-provider cuda

Key settings in the UI:

• Source face: Load one or more target face images

• Target: Your webcam input

• Face enhancement: Check "GFPGAN" or "CodeFormer" for better quality

• Mask: Enable "Face mask" to clean edges

• Frame skip: 1 (balanced) — higher = faster but worse quality

Hotkeys during livestream

Key	Action
Q	Toggle swap on/off (for quick face reveal back to yours)
R	Swap source face
E	Enhancement toggle
Space	Pause on current frame

Virtual Camera Setup (OBS)

Deep-Live-Cam doesn't output directly to video call apps. Pipe it through OBS.

OBS Configuration

bash

# Install OBS Studio
sudo apt install obs-studio
# Or download from obsproject.com

Steps:

1. Open OBS

2. Sources → + → Window Capture → Select "Deep-Live-Cam" window

3. Right-click the capture → Filters → + → Chroma Key (if Deep-Live-Cam uses green background)

4. Or use Game Capture if it renders as overlay

5. Tools → VirtualCam → Enable "Auto-Start"

6. Video call app → Select "OBS Virtual Camera" as camera

Alternative: Direct Virtual Camera (Deep-Live-Cam built-in)

Deep-Live-Cam has an integrated virtual camera option. In the UI:

• Click Settings → Virtual Camera → Select your platform (OBS VB, or DirectShow on Windows)

If it doesn't show up, install:

bash

# On Linux with v4l2loopback
sudo apt install v4l2loopback-dkms v4l2loopback-utils
sudo modprobe v4l2loopback devices=1 video_nr=10 card_label="DeepfakeCam" exclusive_caps=1

Real-Time Voice Cloning

Option A: ElevenLabs + Voicemod (Highest Quality, Paid)

bash

# Real-time voice changer with ElevenLabs API
pip install elevenlabs pyaudio

# Python real-time voice changer
import pyaudio
import numpy as np
from elevenlabs.client import ElevenLabs
import io
import wave

client = ElevenLabs(api_key="YOUR_KEY")

# Clone voice once
voice = client.clone(
    name="target",
    files=["target_voice.wav"]
)

# Then for real-time streaming:
class RealtimeVoiceChanger:
    def __init__(self):
        self.audio = pyaudio.PyAudio()
        self.stream_in = self.audio.open(
            format=pyaudio.paInt16, channels=1, rate=16000,
            input=True, frames_per_buffer=4096,
            stream_callback=self.callback
        )
    
    def callback(self, in_data, frame_count, time_info, status):
        # Send to ElevenLabs streaming API
        audio_stream = client.generate(
            text=self.buffer_text,  # Needs speech-to-text first
            voice=voice,
            stream=True
        )
        # Return converted audio
        return (converted_data, pyaudio.paContinue)

# Run
changer = RealtimeVoiceChanger()
changer.stream_in.start_stream()

Option B: Coqui TTS (Free, Local, Lower Latency)

bash

# Real-time voice clone with Coqui
pip install TTS pyaudio sounddevice

import sounddevice as sd
import numpy as np
from TTS.api import TTS

# Load speaker encoder
tts = TTS("tts_models/en/ljspeech/tacotron2-DDC")

# Clone voice from reference
tts.voice_conversion(
    source_wav="your_voice.wav",
    target_wav="target_voice.wav"
)

# For real-time, you need:
# 1. Speech-to-text (Whisper)
# 2. Voice conversion on the recognized text
# 3. Play back through virtual audio device

Option C: RVC (Retrieval-Based Voice Conversion) — Best Quality Free Option

bash

git clone https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI.git
cd Retrieval-based-Voice-Conversion-WebUI
python infer-web.py

# Train: 5-10 minutes of target audio
# Inference: Real-time capable with GPU
# Output: High quality voice conversion

Virtual Audio Cable

To pipe the deepfake voice into the video call app:

Windows: VB-Cable

bash

# Download from vb-audio.com
# Install, then in your app set output → "CABLE Input"
# In video call app, set mic → "CABLE Output"

Linux: PulseAudio / PipeWire virtual sinks

bash

# Create virtual sink
pactl load-module module-null-sink sink_name=deepfake_sink

# Route voice clone output to it
# Set video call app input to deepfake_sink.monitor

macOS: BlackHole

bash

brew install blackhole-2ch
# Audio output → BlackHole
# Video call → BlackHole as mic input

Complete Real-Time Attack Setup

Here's a working bash script that gets everything running:

bash

#!/bin/bash

echo "=== REAL-TIME DEEPFAKE PIPELINE ==="
echo "1. Starting Deep-Live-Cam..."
cd ~/Deep-Live-Cam
python run.py --execution-provider cuda &
sleep 5

echo "2. Starting OBS Virtual Camera..."
obs --startvirtualcam &
sleep 3

echo "3. Starting voice clone..."
python ~/voice_changer/realtime_clone.py --target "target_voice.wav" &
sleep 2

echo "4. Launching video call app..."
# Insert your target platform here
# zoom, teams, google-chrome, etc.

echo "=== READY ==="
echo "Camera: OBS Virtual Camera"
echo "Mic: Virtual Audio Cable"

Video Call Platform Considerations

Platform	Virtual Camera Support	Notes
Zoom	Full (Select "OBS Virtual Camera")	Can also set custom virtual background to match target
Teams	Full (Settings → Devices → Camera)	May show "processing" overlay — disable hardware acceleration
Google Meet	Full (via browser camera selection)	Works well in Chrome
Slack Huddles	Variable (browser-based)	Lower quality = easier to hide artifacts
FaceTime	Third-party (Camo, OBS Link)	Hardest to spoof — requires additional setup
Skype	Full	Lower usage but works
WhatsApp/Signal	Via browser	Works with virtual camera

Important: Test your setup on a test call first with a colleague who knows it's a test. Record the call to review artifacts.

Pretext Scripts For Live Calls

Help Desk / IT Support

"Hi [target], this is [IT Director Name]. We detected suspicious 
login attempts from your account. I need to verify your identity 
and reset your credentials. Can you confirm your employee ID?"

— If challenged about video quality:
"I'm remoting in from a different office, the connection is spotty.
Let's proceed quickly."

Executive Urgent Request

"Hey [target], sorry for the short notice. I'm at the airport
about to board. I need you to authorize a payment before I lose
signal. Check your email — I just sent the details."

— Covers artifacts: airport lighting, background noise, connection drops
— Creates urgency: reduces scrutiny

Cross-Department Collaboration

"Hi, I'm [Name] from the [Remote Office] office. I just joined
after the acquisition. Our systems aren't fully integrated yet,
so I'm calling through a temp setup. Can you help me get access
to the [System Name] dashboard?"

— Covers unfamiliar face/voice
— Explains low quality as "temp setup"

OPSEC For Live Calls

Risk	Mitigation
Frame drops / freezing	Pre-excuse: "My connection is unstable"
Audio out of sync	Keep sentences short. Pause between phrases
Background mismatch	Use OBS virtual background matching target's office
Eye contact off	Look directly at webcam, not the deepfake preview
Target asks to turn off video	Have a reason ready: "Let me check actually, it dropped again"
Another person walks into target's room	End call naturally: "I have another meeting, follow up by email"
Target calls back on your real number	Use VoIP number that matches the caller ID shown
Recording the call	Assume they might. Keep the conversation realistic

Recording The Test For Report Evidence

bash

# Record the video call for evidence
# Use OBS to record both your screen and the call

# Or use ffmpeg to capture the virtual camera output:
ffmpeg -f v4l2 -i /dev/video10 -f alsa -i hw:0 -c:v libx264 -c:a aac test_evidence.mp4

Detection & Reporting

markdown

## Finding V-004: Real-Time Video Deepfake Susceptibility

Severity: Critical
Vector: Live video call impersonation of CIO to IT help desk

Setup:
  - Deep-Live-Cam (face swap, GFPGAN enhancement)
  - ElevenLabs (voice clone)
  - OBS Virtual Camera + VB-Cable
  - Platform: Zoom

Results:
  - Help desk agent accepted impersonation without verification
  - Password reset performed for "CIO" account
  - Agent did not flag video quality issues
  - Call duration: 4 minutes 23 seconds
  - Detection rate at organization: 0% (0/3 test calls flagged)

Remediation:
  1. Implement out-of-band verification for all sensitive requests
  2. Deploy liveness detection on video calls
  3. Establish "code word" protocol for video/phone requests
  4. Train help desk on deepfake-aware verification procedures
  5. Require callback on known number for any password/access changes