Quantum-Resistant Media Authenticator

Robust audio/video authentication system that survives compression using speech-optimized perceptual hashing and quantum-resistant cryptography.

✨ Key Features

• 🔐 Quantum-Resistant: Ed25519 + future Dilithium support
• 🎙️ Speech-Optimized: Robust features that survive compression
• 📹 Audio + Video: Dual-modal authentication
• 📦 Compression-Resistant: Verified up to 60.9% file reduction
• 🛡️ Tampering Detection: Phase inversion (9.6%), silence (62.4%), noise (46.7%), video (51.0%) - all detected
• ⚡ Two Modes: Embedded (self-contained) or External (.qrmn file)

---

🚀 Quick Start

Installation

pip install -r requirements.txt

Basic Usage

# 1. Generate keys (one time)
python3 keygen.py generate

# 2. Sign media (creates .qrmn file)
python3 external_signer.py example.mp4 example_private.json

# 3. Verify media
python3 external_verifier.py example.mp4 example.mp4.qrmn example_public.json

# 4. Run comprehensive demo
python3 demo.py

Supported Formats

• Video: .mp4, .avi, .mov, .mkv, .webm
• Audio: .wav, .mp3, .flac, .m4a, .ogg

Verification Results

• 🟢 GREEN: Authentic and verified
• 🟡 AMBER: Possibly authentic, verify manually
• 🔴 RED: Not authentic or tampered

---

📊 How It Works

Three-Layer Security

1. File Integrity (SHA-256)

• Cryptographic hash of entire file
• Detects any bit-level modification
• Fast rejection of altered files

2. Content Verification (Perceptual Hashing)

• Audio: F0 patterns, MFCC, spectral envelope, energy, rhythm + polarity fingerprinting
• Video: Edge patterns, color histograms, texture, brightness, contrast
• Compression-resistant features with 80% (audio) / 75% (video) thresholds

3. Cryptographic Chain

• Hash-chain links all segments (SHA-256)
• Ed25519 signatures (quantum-resistant ready)
• Prevents insertion, deletion, reordering

Why Tolerance Instead of Exact Matching?

Compression changes bits but preserves content:

• Original features: [1,0,1,1,0,1,...]
• After compression: [1,0,1,1,1,1,...] (91% similar)
• System accepts ≥80% = authentic ✅

This follows industry best practices from academic research on robust media hashing.

---

🎯 Use Cases

• Speech verification - Podcasts, interviews, recordings
• Deepfake detection - AI-generated content identification
• Legal/forensic evidence - Chain of custody for media
• Content distribution - Verify authenticity across platforms
• Compression tolerance - Works after YouTube, TikTok, messaging apps

---

📈 Performance

Detection Rates

Attack Type	Result	Similarity
Phase Inversion	✅ Detected	9.6%
Silence Injection	✅ Detected	62.4%
Noise Replacement	✅ Detected	46.7%
Video Tampering	✅ Detected	51.0%

Compression Resistance

Scenario	File Reduction	Status
Heavy AAC	-22%	✅ PASS
Extreme Low	61%	✅ PASS
MP3 Transcode	-19%	✅ PASS
OGG Vorbis	22%	✅ PASS
Mono Resample	49%	✅ PASS
Speed Change	27%	✅ PASS

All 6/6 scenarios pass with high confidence (95-100% similarity)

---

🔧 Configuration

Basic Configuration

from src.perceptual_hash import PerceptualHasher

# Balanced (default)
hasher = PerceptualHasher(
    speech_optimized=True,
    polarity_threshold=0.60,    # Detection sensitivity
    polarity_penalty=0.15       # Rejection severity
)

Tuning Options

Conservative (fewer false positives):

polarity_threshold=0.70, polarity_penalty=0.20

Aggressive (maximum security):

polarity_threshold=0.50, polarity_penalty=0.10

---

📊 Technical Details

Audio Perceptual Hashing

Features Extracted (48 bits):

• F0 patterns (pitch)
• MFCC coefficients (vocal tract)
• Spectral envelope (voice quality)
• Energy patterns (amplitude)
• Temporal rhythm (speech rate)

Polarity Fingerprint (32 bits):

• Waveform mean
• Temporal skewness
• Time-weighted features
• Zero-crossing asymmetry
• Peak sign detection

Total: 10 bytes per window

Two-Stage Similarity Check

# Stage 1: Polarity check (fast rejection)
if polarity_similarity < 0.60:
    return full_similarity * 0.15  # 85% penalty
    
# Stage 2: Full comparison
return hamming_similarity(hash1, hash2)

Result:

• Phase inversion: 9.8% similarity (detected)
• Compression: 97.5% similarity (authentic)

Video Perceptual Hashing

Features Extracted (96 bits):

• Edge density (Canny)
• Color histograms (HSV)
• dHash (difference hash)
• Texture analysis
• Brightness/contrast

Total: 12 bytes per keyframe

Hash Chain Structure

Genesis → Window 0 → Window 1 → ... → Window N → Video 0 → Video 1 → ...

Each link contains:

• perceptual_hash: Binary feature vector
• chain_hash: SHA-256(prev_hash + current_features)
• metadata: Timestamp, method, index
• signature: Ed25519 signature

Signature Format (.qrmn files)

{
  "version": "2.0",
  "method": "external",
  "file_hash": "sha256...",
  "media_file": "filename.mp4",
  "total_windows": 55,
  "messages": [...],
  "signatures": [...],
  "audio_info": {...},
  "video_keyframes": [...],
  "video_messages": [...],
  "video_signatures": [...]
}

---

📊 Verification Confidence Levels

Level	Audio	Video	Meaning
🟢 Very High	≥95%	≥90%	Bit-for-bit identical or minimal processing
🟢 High	≥90%	≥80%	Authentic, minor compression
🟢 Medium	≥80%	≥70%	Authentic, heavy compression
🟡 Low	≥70%	≥60%	Possibly authentic, verify manually
🔴 Rejected	<70%	<60%	Not authentic, tampering detected

Why Different Thresholds?

Audio: 80%

• Features (F0, MFCC) very stable under compression
• Achieves 97.5% average → 17.5 point margin

Video: 75%

• Visual features more affected by quantization
• Achieves 89.6% average → 14.6 point margin
• Block artifacts and DCT compression cause more drift

---

⚖️ Limitations

• Embedded mode: Not compression-resistant (use external mode with .qrmn files)
• Tolerance-based: Uses ≥80% similarity threshold, not cryptographic exact matching
• Phase manipulation: Detects full inversion (×-1), but not sophisticated partial phase shifts
• Extreme distortion: May fail with severe noise/distortion beyond tested scenarios

---

🧪 Testing & Validation

# Run rigorous test suite
python3 rigorous_test.py

# Run comprehensive demo (6 compression scenarios)
python3 demo.py

Tests phase inversion, splicing attacks, noise handling, and false positives

What Gets Tested

1. Phase inversion detection (~9.6% similarity)
2. Audio tampering detection (silence, noise, phase inversion)
3. Audio compression resistance (≥80%)
4. Video identity (100% similarity)
5. Video tampering detection (black, white, noise frames)
6. Video compression resistance (≥75%)
7. Hash size validation

---

📁 Project Structure

├── README.md                    # This file
├── requirements.txt             # Python dependencies
├── src/                         # Core modules
│   ├── perceptual_hash.py       # Audio perceptual hashing + polarity detection
│   ├── video_perceptual_hash.py # Video perceptual hashing
│   ├── enhanced_verification.py # Adaptive quality + temporal consistency
│   ├── crypto_signing.py        # Ed25519/Dilithium signatures
│   ├── hashchain.py             # Cryptographic chain
│   ├── canonicalization.py      # Audio preprocessing
│   └── embedding.py             # DCT steganography
├── external_signer.py           # Create .qrmn signature files
├── external_verifier.py         # Verify with .qrmn files
├── embedded_signer.py           # Embed signatures in media
├── embedded_verifier.py         # Verify embedded signatures
├── keygen.py                    # Generate cryptographic keys
├── demo.py                      # Comprehensive demonstration
└── rigorous_test.py             # Rigorous test suite

---

🤝 Contributing

For production use, consider:

• Additional speech feature extraction methods
• More post-quantum algorithm support
• Real-time processing optimization
• GUI interfaces
• API endpoints

---

📄 License

MIT License

---

📚 References

• Ed25519: RFC 8032
• Dilithium: NIST PQC Round 3
• Audio hashing: Robust audio fingerprinting research
• Video hashing: Perceptual hashing literature

---

⚡ Get Started: Run python3 demo.py to see the full system in action!