Deepfake Detection

Deepfake Detection & Media Authentication

Comprehensive framework for detecting synthetic media, analyzing manipulation artifacts, and establishing media provenance in the post-empirical era.

Key Insight: Traditional detection methods (PRNU, IGH, DQ) are like fingerprints—helpful, but disputable. Cryptographic provenance (C2PA) is like a DNA match—mathematically certain (collision probability 2⁻²⁵⁶).

When to Use

• Verifying authenticity of images or videos before publication
• Detecting AI-generated or manipulated media (deepfakes, face swaps, synthetic voices)
• Forensic analysis of suspicious media for legal or journalistic purposes
• Implementing automated media authentication pipelines
• Establishing content provenance and chain of custody
• Countering disinformation campaigns and Advanced Persistent Manipulators (APMs)

Related Skills

• security-audit - Security assessment patterns
• security-incident-reporting - Incident documentation for disinformation attacks
• enterprise-readiness - Infrastructure for automated verification pipelines
• cli-tools - Auto-installation of required tools

1. What Are Deepfakes?

Definition

Deepfakes are synthetic media created using deep learning techniques—primarily Generative Adversarial Networks (GANs), Diffusion Models, and Autoencoders—to generate or manipulate audiovisual content with a high degree of realism. The term combines "deep learning" and "fake."

Types of Synthetic Media

Emerging Capabilities (2025-2026)

The Entertaining Side

Deepfakes have legitimate and creative applications:

Example: The "This Person Does Not Exist" website showcases GAN-generated faces that fascinate users with the uncanny realism of non-existent people.

The Dangerous Side

The same technology enables serious harms:

Real Case (2024): WPP CEO Mark Read was targeted by a sophisticated deepfake voice clone attempting to authorize fraudulent transfers [19]. Deepfake fraud cases surged 1,740% in North America between 2022-2023, with average losses exceeding $500,000 per incident [18].

Current Scale (2025-2026)

Key Insight: The exponential growth rate means detection systems face an ever-increasing volume challenge, reinforcing the need for proactive authentication (C2PA) over reactive detection.

The Future of Deepfakes

The Detection Arms Race

Recent research confirms the growing challenge of detection generalizability [1]:

Generation Quality:    ████████████████████░░░░  85% (2026)
Detection Accuracy:    █████████████░░░░░░░░░░░  55% (2026)
                       ↑ Gap widening over time

Key Insight: We are transitioning from a world where "seeing is believing" to one where "cryptographic proof is believing." The future lies not in perfect detection, but in provenance infrastructure (C2PA v2.3) that proves authenticity at creation [15, 16]. Traditional detection methods (PRNU, IGH, DQ) are like fingerprints—helpful, but disputable. Cryptographic provenance (C2PA) is like a DNA match—mathematically certain.

2. Strategic Context: The Post-Empirical Era

The Crisis of Empirical Evidence (2026)

The boundary between authentic and synthetic media has effectively vanished. Trillion-parameter models have commoditized the generation of photorealistic synthetic content, transforming deepfakes from isolated experiments into an industrialized disinformation capability.

The ABC Framework of Synthetic Media Threats

The 4D Disinformation Tactics

3. System Architecture

LLM Integration Strategy

The skill implements a hierarchical model structure for forensic analysis:

Model Selection Rationale

• Claude Opus 4.5: Best-in-class for nuanced multimodal analysis and synthesizing complex forensic evidence into coherent reports
• Gemini Pro 3.0: Strong visual understanding for cross-validating image/video analysis results
• GLM Pro Thinking 4.7: Chain-of-thought reasoning for transparent forensic logic that can be audited

Architecture Requirements

1. Asynchronous Processing Pipeline: Handle high token counts from multimodal analysis
2. Vector Database for CRF Profiles: Store and query Camera Response Function signatures
3. RAG Integration: Access forensic reference databases during inference
4. Tool Integration: ffmpeg, ExifTool, ImageMagick for low-level signal processing

4. Required Tools & Installation

Tool Overview

Auto-Installation by Agent

When a required tool is missing, the agent will detect this and offer to install it. User approval is required before any installation.

🔧 Tool Missing: ffmpeg

The agent needs 'ffmpeg' for video frame extraction and analysis.
This tool is not currently installed on your system.

Would you like me to install it?
  [macOS]  brew install ffmpeg
  [Ubuntu] sudo apt install ffmpeg
  [Windows] winget install ffmpeg

⚠️ Approval required: Type 'yes' to proceed or 'no' to skip.

Manual Installation

macOS (Homebrew)

# Install all recommended tools
brew install ffmpeg exiftool imagemagick jq

# Optional: C2PA verification tool
brew install c2patool

Ubuntu/Debian

# Install all recommended tools
sudo apt update
sudo apt install ffmpeg libimage-exiftool-perl imagemagick jq

# Optional: C2PA verification tool (from GitHub releases)
curl -L https://github.com/contentauth/c2patool/releases/latest/download/c2patool-linux-x86_64.tar.gz | tar xz
sudo mv c2patool /usr/local/bin/

Windows (winget)

# Install all recommended tools
winget install ffmpeg
winget install exiftool
winget install imagemagick
winget install jqlang.jq

# Optional: C2PA verification tool (from GitHub releases)
# Download from: https://github.com/contentauth/c2patool/releases

Verification

# Verify installations
ffmpeg -version
exiftool -ver
magick -version
jq --version
c2patool --version  # if installed

Tool Usage Examples

ffmpeg for Feature Extraction

# Extract I-frames for PRNU analysis
ffmpeg -i input.mp4 -vf "select='eq(pict_type,I)'" -vsync vfr frame_%04d.png

# Analyze inter-frame consistency (temporal artifacts)
ffmpeg -i input.mp4 -vf "mpdecimate,setpts=N/FRAME_RATE/TB" -c:v libx264 dedup.mp4

# Extract metadata for container audit
ffprobe -v quiet -print_format json -show_format -show_streams input.mp4

# Isolate audio stream for voice clone detection
ffmpeg -i input.mp4 -vn -acodec pcm_s16le -ar 44100 audio.wav

# Extract specific frame range for analysis
ffmpeg -i input.mp4 -ss 00:01:30 -t 00:00:10 -c copy segment.mp4

ExifTool for Metadata Forensics

# Extract all metadata
exiftool -json input.jpg | jq .

# Check for editing software traces
exiftool -Software -CreatorTool -HistorySoftwareAgent input.jpg

# Compare metadata between original and suspected fake
exiftool -g1 -a -u original.jpg suspected.jpg | diff -y - 

# Find GPS coordinates (if present)
exiftool -gps:all -c "%.6f" input.jpg

# Check creation/modification times for inconsistencies
exiftool -time:all -G1 input.jpg

ImageMagick for Image Analysis

# Analyze image statistics (useful for noise analysis)
magick identify -verbose input.jpg

# Extract error level analysis (ELA) for manipulation detection
magick input.jpg -quality 95 ela_temp.jpg
magick composite input.jpg ela_temp.jpg -compose difference ela_output.jpg

# Check for resampling artifacts
magick input.jpg -resize 200% -resize 50% resample_test.jpg

C2PA Tool for Provenance

# Verify C2PA manifest
c2patool verify input.jpg

# Extract manifest details as JSON
c2patool manifest input.jpg -o manifest.json

# Check certificate chain
c2patool trust input.jpg

C2PA Test Files for Validation

Official test files from the C2PA organization (CC BY-SA 4.0):

Source: c2pa-org/public-testfiles

Understanding C2PA Validation: The chain is verified step-by-step: (1) Certificate verified → (2) Signature valid → (3) Claims unchanged → (4) Image hash matches. One failure breaks the entire chain.

5. Forensic Detection Criteria

Criterion A: Sensor Fingerprints (PRNU/PCE)

Photo-Response Non-Uniformity (PRNU) is a sensor-specific noise pattern that acts as a biometric fingerprint for cameras.

Metric: Peak to Correlation Energy (PCE)

# Conceptual PRNU/PCE calculation
def calculate_pce(image: np.ndarray, reference_prnu: np.ndarray) -> float:
    """
    Calculate Peak-to-Correlation-Energy for PRNU matching.
    
    Returns:
        PCE value > 60 indicates high confidence match
        PCE value 40-60 indicates moderate confidence
        PCE value < 40 indicates low confidence or mismatch
    """
    noise_residual = extract_noise_residual(image)
    correlation = correlate_2d(noise_residual, reference_prnu)
    peak = np.max(correlation)
    energy = np.mean(correlation**2)
    return peak**2 / energy

PRNU Limitations (Wronski Effect)

Modern computational photography (multi-frame capture, super-resolution) breaks the direct sensor-to-pixel mapping [22, 23]:

Criterion B: Noise-Intensity Relationship & IGH

The Intensity Gradient Histogram (IGH) classifies the relationship between local intensity and noise, exploiting Camera Response Function (CRF) physics.

def analyze_igh_profile(image: np.ndarray) -> dict:
    """
    Analyze Intensity Gradient Histogram for CRF consistency.
    
    Synthetic blur (portrait mode) destroys the statistical harmony
    of the CRF model, detectable via IGH asymmetry analysis.
    """
    gradients = compute_intensity_gradients(image)
    
    # Authentic optical blur: asymmetric gradient distribution
    # Synthetic blur: symmetric Gaussian distribution
    symmetry_score = measure_gradient_symmetry(gradients)
    
    return {
        "gradient_histogram": gradients,
        "symmetry_score": symmetry_score,
        "classification": "authentic" if symmetry_score < 0.7 else "synthetic"
    }

Criterion C: Geometric & Optical Blur Analysis

Physical law: Due to CRF non-linearity, optically blurred edges are asymmetric. Software Gaussian filters produce symmetric profiles.

Criterion D: Compression Artifacts (Double JPEG / DQ)

Double Quantization (DQ) effects serve as primary evidence for multiple save operations, enabling splicing localization.

def generate_dq_probability_map(image: np.ndarray) -> np.ndarray:
    """
    Generate Double Quantization probability map.
    
    Spliced regions show different quantization histories,
    creating detectable statistical anomalies.
    """
    dct_blocks = compute_dct_blocks(image)
    q_estimates = estimate_quantization_tables(dct_blocks)
    
    # Probability map highlights regions with different
    # compression histories (splicing indicators)
    prob_map = detect_quantization_inconsistencies(q_estimates)
    
    return prob_map  # Heatmap: red = likely manipulation

6. Video-Specific Detection

Video deepfake detection requires analysis of temporal consistency, which current research shows remains challenging for generalization across different manipulation methods [1, 4].

Visual Detection Indicators (Human Review)

Before algorithmic analysis, trained reviewers check for these telltale signs:

Best Practice: Slow down video to 25% speed and examine frame-by-frame. Artifacts become more visible when temporal smoothing is removed.

Temporal Consistency Analysis

def analyze_temporal_artifacts(video_path: str) -> dict:
    """
    Detect temporal inconsistencies in video deepfakes.
    
    Face-swap deepfakes often show:
    - Flickering at face boundaries
    - Inconsistent lighting between frames
    - Unnatural head pose transitions
    """
    frames = extract_frames(video_path)
    
    results = {
        "face_boundary_flickering": detect_boundary_flickering(frames),
        "lighting_consistency": analyze_lighting_consistency(frames),
        "pose_smoothness": measure_pose_transitions(frames),
        "blink_analysis": detect_blink_patterns(frames),  # Early deepfakes lacked blinking
        "audio_visual_sync": check_lip_sync_accuracy(video_path)
    }
    
    return results

GAN & Diffusion Model Fingerprint Detection

Modern detection must address both GAN-based and diffusion model-generated images. Recent research demonstrates that diffusion models leave distinct artifacts detectable via uncertainty estimation [5] and characteristic photorealism patterns [6].

def detect_gan_fingerprints(image: np.ndarray) -> dict:
    """
    Detect characteristic patterns left by generative architectures.
    
    Different families (StyleGAN, Stable Diffusion, DALL-E, Midjourney)
    leave distinct frequency-domain artifacts.
    """
    fft = compute_fft_spectrum(image)
    
    # GANs often produce checkerboard patterns in FFT
    checkerboard_score = detect_checkerboard_artifacts(fft)
    
    # Spectral analysis for GAN-specific signatures
    gan_signatures = match_known_gan_spectra(fft)
    
    return {
        "checkerboard_score": checkerboard_score,
        "suspected_generator": gan_signatures.get("best_match"),
        "confidence": gan_signatures.get("confidence")
    }

7. Semantic Forensics (SemaFor)

When pixel-level artifacts are masked, focus on semantic inconsistencies. This approach was pioneered by the DARPA SemaFor program (2020-2024), which has since transitioned technologies to operational government use [12, 13, 14].

Cross-Modal Consistency Checks

Example: Shadow Analysis

def analyze_shadow_consistency(image: np.ndarray) -> dict:
    """
    Detect physically impossible shadow configurations.
    
    A single light source produces shadows in consistent directions.
    Composited images often have inconsistent shadow angles.
    """
    shadows = detect_shadows(image)
    objects = detect_objects(image)
    
    shadow_vectors = []
    for obj, shadow in zip(objects, shadows):
        vector = compute_shadow_vector(obj, shadow)
        shadow_vectors.append(vector)
    
    # All vectors should converge to consistent light source
    consistency_score = measure_vector_convergence(shadow_vectors)
    
    return {
        "shadow_vectors": shadow_vectors,
        "consistency_score": consistency_score,
        "physically_plausible": consistency_score > 0.85
    }

8. Authenticity Scoring System

Analysis Layer Weighting

Based on scientific reliability of each method (from webconsulting.at forensics research):

Important: Without C2PA verification, maximum achievable grade is 2 ("No manipulation indicators"). Grade 1 ("Provenance cryptographically verified") requires a validated signature chain.

Probability to Grade Mapping

Composite Scoring Algorithm

def calculate_authenticity_score(media_path: str) -> dict:
    """
    Calculate composite authenticity score from multiple forensic signals.
    """
    image = load_media(media_path)
    
    scores = {
        "prnu_pce": analyze_prnu(image),           # Weight: 0.25
        "igh_profile": analyze_igh_profile(image), # Weight: 0.20
        "dq_artifacts": detect_dq_artifacts(image), # Weight: 0.20
        "gan_fingerprints": detect_gan_fingerprints(image), # Weight: 0.15
        "semantic_consistency": check_semantic_consistency(image), # Weight: 0.20
    }
    
    weights = [0.25, 0.20, 0.20, 0.15, 0.20]
    composite = sum(s * w for s, w in zip(scores.values(), weights))
    
    return {
        "authenticity_probability": composite,
        "grade": map_to_grade(composite),
        "detailed_scores": scores,
        "confidence_interval": calculate_confidence(scores)
    }

9. Content Provenance (C2PA / CAI)

C2PA Steering Committee (2026)

The Coalition for Content Provenance and Authenticity is governed by major technology and media companies:

Content Credentials: The CR Icon

C2PA is the technical standard. Content Credentials is the user-facing implementation with the visible "CR" icon.

Key Feature: All assertions are cryptographically signed. Changing even one pixel invalidates the signature—manipulation is immediately detectable.

Industry Adoption (2025-2026)

C2PA is rapidly becoming the industry standard. Current adoption landscape:

Prognosis (3-5 years): For media organizations and government agencies: Without cryptographic provenance, no file will be considered trustworthy.

Content Authenticity Initiative (CAI) Integration

{
  "claim": {
    "recorder": "Canon EOS R5",
    "signature": {
      "alg": "ES256",
      "cert": "-----BEGIN CERTIFICATE-----...",
      "sig": "base64_signature"
    },
    "assertions": [
      {
        "label": "c2pa.actions",
        "data": {
          "actions": [
            {
              "action": "c2pa.created",
              "when": "2026-01-20T14:32:00Z",
              "softwareAgent": "Canon DPP 4.17"
            }
          ]
        }
      }
    ]
  }
}

Provenance Verification Workflow

def verify_c2pa_provenance(media_path: str) -> dict:
    """
    Verify Content Authenticity Initiative (C2PA) manifests.
    
    C2PA provides cryptographic proof of media origin and edit history.
    """
    manifest = extract_c2pa_manifest(media_path)
    
    if not manifest:
        return {
            "has_provenance": False,
            "recommendation": "No cryptographic provenance available. Proceed with forensic analysis."
        }
    
    # Verify certificate chain
    cert_valid = verify_certificate_chain(manifest["signature"]["cert"])
    
    # Verify signature
    sig_valid = verify_signature(
        manifest["claim"],
        manifest["signature"]["sig"],
        manifest["signature"]["alg"]
    )
    
    # Check assertion integrity
    assertions_valid = verify_assertions(manifest["assertions"])
    
    return {
        "has_provenance": True,
        "certificate_valid": cert_valid,
        "signature_valid": sig_valid,
        "assertions_valid": assertions_valid,
        "edit_history": extract_edit_history(manifest),
        "original_device": manifest["claim"].get("recorder"),
        "overall_valid": all([cert_valid, sig_valid, assertions_valid])
    }

10. Forensic Report Template

Module A: Media Metadata & Summary

# Media Authentication Report

## Metadata
| Field | Value |
|-------|-------|
| Report ID | MAR-2026-001 |
| Classification | Confidential |
| Analysis Date | 2026-01-23 15:00 UTC |
| Media Type | Video (MP4) |
| Duration | 00:02:34 |
| Resolution | 1920x1080 |
| File Hash (SHA-256) | a1b2c3d4... |
| Lead Analyst | Forensic AI Agent |

## Executive Summary (max 200 words)

Analysis of [MEDIA_FILE] reveals [AUTHENTICITY_ASSESSMENT].
Key findings include [PRIMARY_INDICATORS]. The composite authenticity
score is [SCORE]% (Grade: [GRADE]). [RECOMMENDATION].

## Authenticity Assessment
| Criterion | Score | Status |
|-----------|-------|--------|
| PRNU/PCE Match | 45/100 | ⚠️ Inconclusive |
| IGH Profile | 82/100 | ✅ Consistent |
| DQ Artifacts | 23/100 | ❌ Detected |
| GAN Fingerprints | 15/100 | ❌ Detected |
| Semantic Consistency | 67/100 | ⚠️ Minor issues |
| **Composite Score** | **34%** | **Grade: 5** |

Module B: Technical Evidence

## Technical Analysis

### PRNU/PCE Analysis
- Reference device: Not identified
- PCE Value: 23.4 (below threshold of 40)
- Interpretation: Cannot establish camera origin

### Double Quantization Map
![DQ Probability Map](dq_map.png)
- Red regions indicate areas with different compression histories
- Face region shows 89% probability of splicing

### GAN Fingerprint Analysis
- Checkerboard artifacts: Detected (FFT analysis)
- Suspected generator: StyleGAN2-derived architecture
- Confidence: 78%

### Semantic Inconsistencies
1. Shadow direction: Inconsistent (2 apparent light sources)
2. Eye reflections: Different scene reflected in each eye
3. Audio-visual sync: 120ms average desync detected

11. Defense Strategies

Technical Controls

Organizational Inoculation

1. Pre-bunking: Educate stakeholders about deepfake capabilities before exposure
2. Source Triangulation: Verify claims through multiple independent sources
3. Temporal Delay: Wait for verification before amplifying uncertain content
4. Provenance Requirement: Mandate C2PA for critical communications

Response When Targeted (Personal Deepfakes)

If you or your client are depicted in a deepfake:

Legal Framework (Austria)

Austrian Resources:

• Saferinternet.at Helpline - Expert counseling
• Saferinternet.at Unterrichtsmaterialien - Teaching materials ("Wahr oder falsch im Internet")
• Rat auf Draht: 147 (24/7 hotline for young people)
• Internet Ombudsstelle: ombudsstelle.at

Note: Similar protections exist across EU member states under the Digital Services Act (DSA) and GDPR. Consult local counsel for jurisdiction-specific advice.

Detection Pipeline Example

class MediaAuthenticationPipeline:
    """
    Production pipeline for automated media authentication.
    """
    
    def __init__(self, config: PipelineConfig):
        self.prnu_analyzer = PRNUAnalyzer(config.prnu_db)
        self.igh_classifier = IGHClassifier(config.igh_model)
        self.dq_detector = DQDetector()
        self.gan_detector = GANFingerprintDetector(config.gan_signatures)
        self.semantic_analyzer = SemanticAnalyzer(config.llm_endpoint)
        self.c2pa_validator = C2PAValidator(config.trusted_roots)
    
    async def authenticate(self, media_path: str) -> AuthenticationResult:
        # Check provenance first (fast path)
        provenance = await self.c2pa_validator.verify(media_path)
        if provenance.valid:
            return AuthenticationResult(
                authentic=True,
                confidence=0.99,
                method="cryptographic_provenance"
            )
        
        # Run forensic analysis in parallel
        results = await asyncio.gather(
            self.prnu_analyzer.analyze(media_path),
            self.igh_classifier.classify(media_path),
            self.dq_detector.detect(media_path),
            self.gan_detector.detect(media_path),
            self.semantic_analyzer.analyze(media_path)
        )
        
        # Fuse signals
        composite = self.fuse_signals(results)
        
        return AuthenticationResult(
            authentic=composite.score > 0.75,
            confidence=composite.confidence,
            grade=composite.grade,
            details=results,
            method="forensic_analysis"
        )

12. Tool & Dataset References

Detection Tools

Reference Datasets

Industry Standards

• C2PA (Coalition for Content Provenance and Authenticity): Cryptographic media provenance
• CAI (Content Authenticity Initiative): Adobe-led provenance standard
• IPTC Photo Metadata: Standard metadata for photographic content

13. Checklists

Pre-Analysis Checklist

• Obtain original file (avoid screenshots, re-uploads)
• Preserve file hash (SHA-256) for chain of custody
• Document source and context of media
• Check for C2PA/CAI provenance data
• Identify claimed device/source for PRNU matching

Analysis Checklist

• Run PRNU/PCE analysis (if reference available)
• Generate IGH profile and classify blur types
• Create DQ probability map for splicing detection
• Analyze for GAN fingerprints (FFT spectrum)
• Check semantic consistency (shadows, reflections, physics)
• For video: temporal consistency, audio-visual sync
• Calculate composite authenticity score

Reporting Checklist

• Document all findings with confidence levels
• Include visualizations (DQ maps, FFT spectra)
• Provide grade interpretation with caveats
• List limitations of analysis
• Recommend human review for borderline cases

14. Limitations & Caveats

Known Detection Challenges

C2PA Challenge: Screenshots and social media uploads can strip metadata ("stripping attack"). The industry is developing invisible watermarks that survive re-encoding and link back to C2PA manifests.

Ethical Considerations

1. False Positives: Incorrectly flagging authentic media can cause harm
2. Dual Use: Detection research enables better synthesis
3. Automation Bias: Over-reliance on automated verdicts
4. Privacy: PRNU databases can identify individuals

The Liar's Dividend

The mere existence of deepfakes allows bad actors to dismiss authentic evidence as fake. Detection tools must be communicated carefully to avoid amplifying this effect.

References

Academic Research (2024-2025)

1. Ramanaharan, R. et al. (2025). "DeepFake video detection: Insights into model generalisation." Forensic Science International: Digital Investigation, Vol. 52. DOI: 10.1016/j.fsidi.2025.301875

   • Systematic review of generalizability in deepfake detection techniques

2. Ahmed, N. et al. (2024). "Visual Deepfake Detection: Review of Techniques, Tools, and Datasets." IEEE Access, Vol. 12, pp. 180234-180261. DOI: 10.1109/ACCESS.2024.3511641

   • Comprehensive review covering 2018-2024 with 16+ citations

3. Cassia, M. et al. (2025). "Deepfake Forensic Analysis: Source Dataset Attribution and Legal Implications." arXiv:2505.11110. arXiv

   • Dataset attribution for legal proceedings

4. Nature Scientific Reports (2025). "Deepfake video deception detection using visual attention mechanisms." *Sci R

…(truncated)