MIID (Multimodal Inorganic Identity Dataset) is a next-generation identity testing and identity data generation subnet designed to enhance fraud detection, KYC systems, and biometric verification. Our goal is to provide financial institutions, security systems, and AI researchers with a robust dataset of identity-preserving face image variations that help identify deepfake and presentation-attack evasion techniques.
By incentivizing miners to create high-quality face image variations, MIID serves as a critical tool in financial crime prevention, identity resolution, and security intelligence.
Fraudsters use identity manipulation techniques to evade detection — including deepfakes, screen replays, and biometric spoofing. Sanctioned individuals, high-risk entities, and money launderers exploit weaknesses in KYC and IDV systems.
MIID tests and enhances these systems by:
- ✅ Simulating Face-Based Adversarial Scenarios for KYC and biometric screening
- ✅ Evaluating Identity-Preserving Image Transformations
- ✅ Providing Adversarial Face Data for Model Training
This network helps governments, financial institutions, and researchers improve their fraud detection models, making the financial ecosystem safer.
Miners process image variation requests from validators and return identity-preserving face image variations.
- Receive base face images and variation requirements from validators
- Generate variations using diffusion models: pose_edit, lighting_edit, expression_edit, background_edit, and screen_replay (Cycle 2)
- Encrypt and upload results to S3; return signed submission references
- Image generation is the only scored task — a GPU and the full image stack are required to earn rewards
Validators ensure the dataset maintains high-quality and real-world relevance.
- Issue face image variation challenges
- Run automated pre-checks and identity preservation validation
- Perform manual validation to assess transformation accuracy
- Set miner weights based on image variation quality and reputation
- Python 3.10+
- GPU with 8GB+ VRAM (NVIDIA CUDA or Apple Silicon MPS)
- Hugging Face account and API token (for diffusion model access)
- Bittensor wallet with TAO
- 16GB+ RAM (32GB recommended)
- 80GB+ free disk (for model weights and cache)
- Open port 8091 for miner-to-validator communication (Network Setup Guide)
# Install dependencies (includes image-generation stack)
bash scripts/miner/setup.sh
# Activate the miner environment
source miner_env/bin/activate
# Set required environment variables
export HF_TOKEN="hf_YOUR_TOKEN_HERE"
export FLUX_DEVICE="cuda" # or mps for Apple Silicon
# Start mining
pm2 start python --name neuron-miner -- neurons/miner.py --netuid 54 --wallet.name your-wallet --wallet.hotkey your-hotkey --subtensor.network finney# Install dependencies
bash scripts/validator/setup.sh
# Activate the validator environment
source validator_env/bin/activate
# Start validating
pm2 start python --name neuron-validator -- neurons/validator.py --netuid 54 --wallet.name your_wallet --wallet.hotkey your_hotkey --subtensor.network finneyFor detailed instructions, check our Mining Guide and Validator Guide.
- Help banks, fintech, and law enforcement agencies strengthen KYC and biometric fraud detection.
- Contribute to privacy-preserving AI research.
- Earn rewards while enhancing AI-driven face verification and presentation-attack detection.
- Miners: Earn rewards for producing high-quality, identity-preserving face image variations.
- Validators: Gain influence in network security and reward distribution.
MIID is not just another AI dataset—it's a live, evolving system that challenges and improves real-world fraud detection models. Every contribution makes financial systems safer and more secure.
Phase 1: Initial Launch & Name-Based Threat Scenarios (June 2025) Read more details here
- Deploy MIID subnet on Bittensor mainnet.
- Enable validators to test known threat scenarios against miner responses.
- Introduce name-based execution vectors: phonetic, orthographic, and rule-based variations.
- Expand Threat Scenario Query System to allow miners to propose unknown threat scenarios.
- Introduce a Post-Evaluation System to systematically validate and assess new miner-submitted threat scenarios.
- Support new evasion tactics, including nickname-based threats, transliteration-based alterations, and middle name manipulations.
- Improve validator scoring and introduce penalties for repetitive or low-value submissions.
- Add support for location-based unknown attack vectors (UAV) and obfuscation patterns.
- Establish post-validation workflows and LDS V1 (beta → full) to separate signal from noise.
- Use validated UAV quality to build a reputation signal that carries into future cycles.
- Validator-provided seed face images and deepfake-style transformation families.
- Cycle 1: pose_edit, lighting_edit, expression_edit, background_edit. Cycle 2 adds screen_replay.
- Image generation is the sole scored miner task in the current cycle.
- Expand biometric attack families beyond Cycle 1 (e.g., swap/recapture/morphing) (Q1 2026)
- Generate and validate synthetic documents (Q2 2026)
- Simulate digital presence and interactions (Q3 2026)
- Introduce financial transaction modeling (Q4 2026)
- Build 3D identity avatars (Q2 2027)
- Add voice and conversational AI support
- Train a comprehensive model for identity screening.
- Launch a decentralized platform for collaborative validation and contribution.
We are continuously improving MIID to:
- Expand face-based adversarial data generation for enhanced AI benchmarking.
- Integrate more complex biometric attack families (document spoofing, voice, 3D avatars).
- Improve fraud detection AI using multi-modal data sources.
Join us in shaping the future of identity verification and fraud prevention.
📢 Follow the project & contribute to our open-source development!
Discord | GitHub
This project is licensed under the MIT License - see the LICENSE file for details.
Built with ❤️ by the YANEZ-MIID Team