Detection Techniques

Manual Detection Methods

Visual Analysis Checklist

□ Check eye reflections (should match light sources)
□ Observe blinking patterns (natural vs. robotic)
□ Examine face boundaries (blurring, artifacts)
□ Verify skin texture consistency
□ Look for lighting mismatches
□ Check hair movement realism
□ Analyze facial expressions
□ Verify lip-sync accuracy
□ Check for temporal inconsistencies
□ Examine background stability

Audio Analysis

□ Listen for robotic cadence
□ Check background noise consistency
□ Verify breathing patterns
□ Analyze emotional tone authenticity
□ Compare to known voice samples
□ Check for audio artifacts
□ Verify speech patterns
□ Analyze prosody (intonation, stress, rhythm)

Automated Detection Tools

Open Source Solutions

1. Deepware Scanner - Browser-based detection

   • URL: https://scanner.deepware.ai
   • Accuracy: ~75%
   • Free to use

2. Sensity - Video verification platform

   • Real-time analysis
   • API available
   • Enterprise support

3. FaceForensics++ - Research benchmark

   • 1.8M+ images
   • Multiple detection methods
   • Academic use

Commercial Solutions

1. Intel FakeCatcher - Real-time detection

   • 96% accuracy rate
   • Blood flow analysis
   • Enterprise deployment

2. Microsoft Video Authenticator

   • Confidence scores
   • Frame-by-frame analysis
   • Integration with Office 365

3. Truepic - Media authentication

   • Blockchain verification
   • Chain of custody
   • Legal admissibility

Source: Tolosana et al., 2020 - DeepFakes and Beyond: A Survey

Technical Detection Methods

Metadata Analysis

# Check video metadata
exiftool video.mp4 | grep -i "create\|modify\|software"

# Verify file integrity
ffmpeg -i video.mp4 -f null -

# Check for compression artifacts
ffprobe -v error -select_streams v:0 -show_entries stream=codec_name,width,height,r_frame_rate video.mp4

Frame-by-Frame Analysis

import cv2
import numpy as np

def analyze_frames(video_path):
    cap = cv2.VideoCapture(video_path)
    inconsistencies = []
    
    while cap.isOpened():
        ret, frame = cap.read()
        if not ret:
            break
            
        # Check for artifacts and anomalies
        if detect_artifacts(frame):
            frame_num = cap.get(cv2.CAP_PROP_POS_FRAMES)
            inconsistencies.append(frame_num)
    
    cap.release()
    return inconsistencies

def detect_artifacts(frame):
    # Check for common deepfake artifacts
    # - Unnatural color transitions
    # - Blurring at face boundaries
    # - Inconsistent lighting
    return False  # Placeholder

Forensic Analysis Approaches

Spatial Analysis:

• CNN-based face detection
• Facial landmark analysis
• Texture inconsistency detection

Temporal Analysis:

• Optical flow analysis
• Frame-to-frame consistency
• Biological signal detection (blood flow)

Frequency Domain:

• Fourier analysis
• Wavelet decomposition
• Spectral anomaly detection

Source: Rossler et al., 2019 - FaceForensics++

Verification Strategies

Multi-Source Verification

1. Cross-reference with official sources
2. Reverse image search for original content
3. Contact verification - Reach out directly
4. Timestamp analysis - Check publication dates
5. Source credibility - Verify publisher

Context Clues

• Does the content match known behavior?
• Is the source credible and verifiable?
• Are there other versions available?
• What’s the motivation for sharing?
• Does the timing seem suspicious?

Detection Accuracy Comparison

Red Flags & Warning Signs

High-Risk Scenarios

⚠️ Urgent financial requests ⚠️ Sensitive information requests ⚠️ Out-of-character behavior ⚠️ Unusual communication channels ⚠️ Pressure for immediate action ⚠️ Requests for secrecy ⚠️ Unusual emotional state

Technical Red Flags

⚠️ Unnatural eye movements ⚠️ Inconsistent lighting ⚠️ Blurring at face boundaries ⚠️ Unnatural blinking patterns ⚠️ Audio-visual misalignment ⚠️ Background inconsistencies

Statistics

• 96% of deepfakes are non-consensual content
• 500% increase in deepfake incidents (2022-2024)
• $250M+ in documented fraud losses
• $243K average incident cost in financial sector

Source: Sensity AI - State of Deepfakes Report

Next: Prevention Strategies →