🎵 Melodic Metadata Massacrer (MMM)

"In the symphony of digital rights, we are the conductors of chaos." 🎼⚡

MMM is a Python CLI tool that strips metadata, disrupts watermark patterns, and applies spectral perturbation to MP3 and WAV audio files, making it harder for AI-detection systems to identify them as machine-generated.

Note: Sanitization re-encodes audio (MP3 at 320kbps, WAV at PCM_16) and applies subtle spectral modifications. This is not lossless — audio quality is preserved but not bit-identical.

🎭 Features

Core Capabilities

Complete Metadata Annihilation: Removes ID3, RIFF INFO, FLAC tags, and custom chunks
AI Watermark Detection: Identifies spread spectrum, echo-based, and statistical watermarks
Spectral Cleaning: Advanced frequency-domain watermark removal
Fingerprint Elimination: Normalizes AI-generated statistical patterns
Paranoid Mode: Maximum destruction with multiple cleaning passes
Batch Processing: Parallel processing of entire directories
Verification Engine: Before/after comparison with forensic reporting
Web Interface: Local drag-and-drop browser UI via mmm server

Detection Methods

Spread spectrum watermarks
Echo-based signatures
Statistical pattern analysis
Phase modulation detection
Amplitude modulation analysis
Frequency domain anomalies

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/geeknik/mmm.git
cd mmm

# Create virtual environment (Python 3.9+)
python3 -m venv mmm_env
source mmm_env/bin/activate  # On Windows: mmm_env\Scripts\activate

# Install the package and all dependencies
pip install -e .

GPU Acceleration (Optional)

For faster analysis on systems with NVIDIA GPUs:

# Install GPU acceleration packages (NVIDIA GPU required)
pip install cupy-cuda12x torch torchaudio

# Verify GPU detection
python -c "import torch; print(f'CUDA: {torch.cuda.is_available()}')"

GPU Requirements:

NVIDIA GPU with CUDA support
4GB+ VRAM recommended
CUDA 12.x compatible drivers

Performance

Processing speed depends on hardware, file length, and mode:

| Mode | 3.5 min MP3 (4.9 MB) | Notes | |------|----------------------|-------| | Turbo (CPU) | ~70s | 3x real-time, --turbo flag | | Turbo + Paranoid (CPU) | ~99s | 2.2x real-time, all stealth flags | | Turbo (GPU) | Significantly faster | Requires NVIDIA GPU + CUDA | | Regular (no turbo) | Very slow | Runs 6 O(N) watermark detection methods; not recommended for files > 1 min |

Basic Usage

# Sanitize a single file (recommended: turbo + paranoid)
mmm obliterate music.mp3 --turbo --paranoid -o clean_music.mp3

# Quick sanitize without paranoid
mmm obliterate music.mp3 --turbo -o clean_music.mp3

# Batch process directory
mmm massacre /path/to/music --paranoid --workers 8

# Analyze file without modifying
mmm analyze music.mp3 --turbo

🔧 Configuration

MMM uses YAML configuration files for customization:

# ~/.config/mmm/config.yaml
paranoia_level: medium
preserve_quality: high
watermark_detection:
  - spread_spectrum
  - echo_based
  - statistical
output_format: preserve
backup_originals: true

Presets

stealth: Maximum paranoia, quality preservation
stealth-plus: Stealth with advanced flags optimized for detector evasion
fast: Quick processing, basic cleaning
quality: Preserve maximum audio quality
research: Deep analysis, detailed logging

# Use preset
mmm config preset stealth

# Create custom preset
mmm config create my_preset --paranoid maximum --quality high

🎯 Commands

`obliterate`

Complete sanitization of individual files

mmm obliterate INPUT_FILE [OPTIONS]

Options:
  -o, --output PATH     Output file path
  --paranoid           Maximum destruction mode
  --verify             Verify watermark removal
  --backup             Create backup of original
  --format FORMAT      Output format (preserve/mp3/wav)
  --turbo              Enable turbo mode (faster, uses preserving sanitizer)

`massacre`

Batch processing of directories

mmm massacre DIRECTORY [OPTIONS]

Options:
  -d, --output-dir PATH  Output directory
  -e, --extension TEXT   File extensions (multiple)
  -w, --workers INT      Parallel workers
  --paranoid            Paranoid mode
  --backup              Create backups

`analyze`

Forensic analysis without modification

mmm analyze INPUT_FILE                # Regular mode (slow on long files)
mmm analyze INPUT_FILE --turbo        # Turbo mode (faster)

`server`

Browser-based drag-and-drop interface

mmm server                              # Start on localhost:8778
mmm server --port 9000                  # Custom port
mmm server --host 0.0.0.0              # Network-accessible (warns)
mmm server --max-size 1000             # 1 GB upload limit

Options:
  --host TEXT         Bind address  [default: 127.0.0.1]
  --port INTEGER      Listen port  [default: 8778]
  --max-size INTEGER  Max upload in MB  [default: 500]

Open http://127.0.0.1:8778 in your browser to drag-and-drop audio files for sanitization. Supports MP3, WAV, and FLAC. Processing uses the turbo/preserving sanitizer path with optional paranoid mode.

`config`

Configuration management

mmm config              Show current config
mmm config preset NAME  Apply preset
mmm config list         List available presets
mmm config create NAME  Create custom preset
mmm config delete NAME  Delete custom preset
mmm config reset        Reset to defaults

🎛️ Advanced Stealth Flags

These are opt-in, fine-grained toggles for research tuning. Defaults keep audio quality high; enable selectively:

--gated-resample-nudge/--no-gated-resample-nudge (default off): ultra-tiny resample up/down applied only on higher-energy segments (minimal audibility, good stealth).
--phase-noise/--no-phase-noise (default on): tiny FFT phase noise.
--phase-swirl/--no-phase-swirl (default on): light all-pass swirl.
--phase-dither/--no-phase-dither (default on), --comb-mask/--no-comb-mask, --transient-shift/--no-transient-shift: earlier experimental steps (may affect audio; leave off unless testing).
--masked-hf-phase/--no-masked-hf-phase (default off): HF-only masked phase noise.
--micro-eq-flutter/--no-micro-eq-flutter (default off): RMS-gated, <0.013 dB band flutter.
--hf-decorrelate/--no-hf-decorrelate (default off): decorrelate only 12–16 kHz band.
--refined-transient/--no-refined-transient (default off): ultra-small, onset-gated shifts.
--adaptive-transient/--no-adaptive-transient (default off): onset-strength adaptive micro-shifts (~0.03–0.08 ms) with light blending.

Maximum stealth (all flags enabled)

mmm obliterate input.mp3 -o output.mp3 --turbo --paranoid \
  --masked-hf-phase --gated-resample-nudge --micro-eq-flutter \
  --hf-decorrelate --adaptive-transient

Preset shortcut

mmm config preset stealth-plus

Preset stealth-plus includes advanced flags:

phase_dither=False, comb_mask=False, transient_shift=False
phase_swirl=False, masked_hf_phase=False, resample_nudge=False
gated_resample_nudge=True, phase_noise=True
micro_eq_flutter=False, hf_decorrelate=False
refined_transient=False, adaptive_transient=False

Notes on pattern suppression counts:

"Patterns Found/Suppressed" in sanitization results come from the spectral cleaner's suppression actions (e.g., attenuating suspicious bands/patterns) and do not imply detector-verified watermarks unless the detector reports them.
Verification threat counts include metadata/container anomalies and detector findings; if the detector reports zero watermarks, remaining threats are likely metadata/binary anomalies rather than confirmed watermarks.

🧪 Detector Notes (Research)

We test against third-party detectors to understand robustness (not to guarantee evasion). Results on a Suno-generated 3.5 min MP3 (March 2026):

SubmitHub / SHLabs results:

| Mode | Verdict | Spectral: Human | Spectral: Pure AI | Temporal: Human | Temporal: Pure AI | |------|---------|-----------------|-------------------|-----------------|-------------------| | Turbo (default flags) | Possible AI Detected | 15% | 36% | 44% | 6% | | Turbo + Paranoid + All Flags | Inconclusive | 43% | 15% | 65% (likely) | 1% (highly unlikely) |

Paranoid mode with all stealth flags shifted the verdict from "Possible AI Detected" to "Inconclusive"
Temporal: Pure AI dropped to 1% ("highly unlikely")
Aggressive stacks (phase dither / comb mask / transient shift) degraded audio; not recommended

Always audition audio locally before running external checks.

🛡️ Legal & Ethical Notice

⚠️ IMPORTANT: This tool is designed exclusively for authorized security research and educational purposes.

Use only on files you own or have explicit permission to modify
You are responsible for compliance with applicable laws and terms of service
The developers do not condone or support copyright infringement
This tool demonstrates vulnerabilities in watermarking systems for research purposes

📊 Technical Details

Architecture

┌─────────────────────────────────────────┐
│                CLI Layer                │
│  Click-based interface with personality  │
└─────────────────┬───────────────────────┘
                  │
┌─────────────────▼───────────────────────┐
│            Core Processing              │
│  • AudioSanitizer main engine          │
│  • PreservingSanitizer (turbo mode)    │
│  • FileProcessor for batch operations   │
└──

Mmm

Install / Use

README