Deepfake detection and analysis

Deepfake Detection & Analysis

When to Use

• When verifying the authenticity of sensitive media (executives, political figures)
• During incident response for CEO impersonation / BEC scams using voice cloning
• When investigating disinformation or influence operations
• When training corporate teams on deepfake recognition
• When building defensive pipelines for media upload portals

Prerequisites

• Access to target AI/ML system or local model deployment for testing
• Python 3.9+ with relevant ML libraries (transformers, torch, openai)
• Understanding of LLM architecture and prompt processing pipelines
• Authorized scope and rules of engagement for AI red team testing

Workflow

Phase 1: Visual Artifact Inspection (Manual)

# Extract frames from video for detailed analysis
ffmpeg -i suspicious_video.mp4 -vf fps=1/1 out%04d.png

# Look for these common visual artifacts:
# 1. Blinking anomalies (too much, too little, unnatural eyelids)
# 2. Lighting/Shadow inconsistencies (face lighting differs from background)
# 3. Blurring or edge artifacts around the face perimeter (the "mask" line)
# 4. Unnatural teeth rendering (AI struggles with individual teeth)
# 5. Mismatched skin tones or unnatural smoothness
# 6. Glitches during rapid movement or hand occlusion (hands passing in front of face)
# 7. Asymmetrical reflections in the eyes

Phase 2: Metadata & Provenance Analysis

# Check EXIF and metadata for manipulation traces
exiftool suspicious_media.jpg
exiftool suspicious_video.mp4

# Look for:
# - Missing standard camera metadata (Make, Model)
# - Software signatures (e.g., Photoshop, AfterEffects, Stable Diffusion)
# - Mismatched timestamps (creation vs. modification)
# - Missing or altered GPS data

# Check for C2PA (Coalition for Content Provenance and Authenticity) manifests
# Many legitimate AI tools now embed watermarks or provenance data

# Search for the source media (Reverse Image Search)
# Provide the isolated face or background to Yandex/Google images
# to find the original unaltered source material

Phase 3: Automated Image/Video Detection

# Use Python and OpenCV/Dlib/PyTorch for automated detection

import cv2
import dlib
import numpy as np
# from deepfake_models import load_detector (example API)

def extract_face_landmarks(image_path):
    """Analyze facial landmarks for unnatural positioning or jitter over time."""
    detector = dlib.get_frontal_face_detector()
    predictor = dlib.shape_predictor('shape_predictor_68_face_landmarks.dat')
    
    img = cv2.imread(image_path)
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    faces = detector(gray)
    
    for face in faces:
        landmarks = predictor(gray, face)
        # Analyze symmetry and proportions
        # Deepfakes often disrupt natural biometric ratios

# Frequency domain analysis
# Deepfakes often lack high-frequency details present in real images
def frequency_analysis(image_path):
    img = cv2.imread(image_path, 0)
    f = np.fft.fft2(img)
    fshift = np.fft.fftshift(f)
    magnitude_spectrum = 20*np.log(np.abs(fshift))
    # Analyze magnitude spectrum for distinct GAN/diffusion artifacts

Phase 4: Voice Clone (Audio) Analysis

# Audio analysis is critical for modern vishing attacks

import librosa
import numpy as np

def analyze_audio_artifacts(audio_path):
    # Load audio
    y, sr = librosa.load(audio_path)
    
    # 1. Check for unnatural breathing patterns
    # AI voices often lack breaths, or place them unnaturally
    
    # 2. Spectrogram analysis
    # Look for frequency cutoffs (synthetic voices often drop frequencies above 10kHz)
    S = librosa.feature.melspectrogram(y=y, sr=sr)
    
    # 3. Analyze pacing and emotion
    # Does the emotion match the semantic content?
    
    # 4. Check for robotic metallic ringing (vocoder artifacts)
    # Common in older or fast-generated voice clones

# Use commercial/open-source detection APIs
# e.g., Resemble AI Detect, AI Voice Detector

Phase 5: Assessing the Delivery Mechanism

# Deepfakes are rarely delivered in a vacuum. Analyze the attack vector:
# 1. Did it come via an urgent WhatsApp/Signal message?
# 2. Is the caller ID spoofed? (Check telecom routing logs if available)
# 3. Did the deepfake video call happen over Zoom/Teams with a "broken camera" excuse?
# 4. Analyze email headers if sent via email

🔵 Blue Team Detection & Defense

• Liveness Checks: Require users to turn their head side-to-side or pass hands over their face during live video verification (breaks facial tracking).
• Challenge-Response: During suspected voice clones, ask a question only the real person would know (that isn't public on social media).
• Watermarking: Implement invisible watermarking for internal corporate media to establish baseline provenance.
• Verification Protocols: Implement strict callbacks for any financial requests, regardless of who appears to be calling.

Key Concepts

Concept	Description
Deepfake	Synthetic media where a person in existing image/video is replaced by someone else
Voice Cloning	Synthesizing a person's voice from a short audio sample
C2PA	Open technical standard providing publishers/creators a way to opt-in to tracing origin
Spectral Analysis	Analyzing the frequencies of an image/audio to find synthetic generation artifacts
Vishing	Voice phishing, increasingly augmented with real-time AI voice changers

Output Format

Synthetic Media Analysis Report
===============================
Subject: Urgent Wire Transfer Voicemail (CEO Impersonation)
Date of Analysis: 2024-X-X
Conclusion: HIGH PROBABILITY SYNTHETIC AUDIO (Voice Clone)

Artifacts Detected:
1. Spectrogram Analysis: Hard frequency cutoff at 11kHz, typical of ElevenLabs generation.
2. Acoustic Anomalies: Complete absence of inhalation/breathing sounds over a 45-second clip.
3. Metadata: Audio file stripped of all recording device metadata, encoded via standard FFmpeg without hardware signatures.
4. Delivery Vector: Received via VoIP number matching known spoofing patterns, not the CEO's registered mobile device.

Recommendation: DO NOT process the wire transfer. Initiate internal incident response for targeted social engineering.

🛡️ Remediation & Mitigation Strategy

• Input Validation: Sanitize and strictly type-check all inputs.
• Least Privilege: Constrain component execution bounds.

📚 Shared Resources

For cross-cutting methodology applicable to all vulnerability classes, see:

• _shared/references/elite-chaining-strategy.md — Exploit chaining methodology and high-payout chain patterns
• _shared/references/elite-report-writing.md — HackerOne-optimized report writing, CWE quick reference
• _shared/references/real-world-bounties.md — Verified disclosed bounties by vulnerability class

References

• MITRE ATLAS: AML.T0048 - Synthetic Media
• Deepware Scanner: Deepware
• Reality Defender: Reality Defender