How are detection models evaluated in GAS?

Discriminative miners submit models in safetensors format. Each model passes an entrance exam (80% accuracy threshold on a reduced dataset within 1 hour 25 minutes) before being benchmarked against the full dataset suite, which includes public benchmarks, private holdout datasets, and weekly-refreshed synthetic media from GAS-Station. Models are scored using `sn34_score`, a geometric mean of normalized MCC (classification accuracy) and Brier score (calibration quality), weighted across multiclass provenance labels (real, synthetic, semisynthetic, and rendered for video). One model per hotkey per modality; new models require new hotkeys.

What hardware do discriminative miners need?

Discriminative miners do not need to host inference hardware. Models are submitted as files and evaluated on validator infrastructure. Generative miners, by contrast, run FastAPI servers on their own hardware and require GPUs (or external API access) to generate synthetic media on demand in response to validator prompts.

How do generative miners earn rewards?

Generative miners earn base rewards proportional to their validation pass rate (content that passes C2PA signature checks and prompt alignment validation) and a multiplier based on fool rate—how often their synthetic media causes discriminators to misclassify. The multiplier also includes sample-size bonuses (logarithmic scaling up to 2× for high evaluation volume). Rewards are weighted by the cost of the generation model used; more expensive models earn higher per-sample rewards, incentivizing investment in quality infrastructure.

What does C2PA verification mean for generative miners?

C2PA (Coalition for Content Provenance and Authenticity) is a metadata standard that digitally signs synthetic media to prove its origin. Validators reject unsigned content from generative miners to prevent gaming and wasted API costs. Supported generation services include OpenAI (DALL-E, Sora), OpenRouter (Google Veo, ByteDance Seedance), Stability AI, and Runway—all of which produce C2PA-signed output. Miners can verify local files with `gascli generator verify-c2pa `.

How is the reward split among detection and generation?

Total emissions are split: 40% image detection lane, 40% video detection lane, 4% audio detection lane, 16% generators. Within each detection lane, emissions follow a 'King of the Hill' model: 85% to the current best model, 10% to the previous best, 5% to the model ranked third. A challenger must score at least 0.01 higher than the reigning model to take the lane. When a new benchmark version releases, the defending king must re-evaluate to keep the crown (dethroning is frozen during defense).

What modalities and classification classes are supported?

Image models classify into three classes: real, synthetic, and semisynthetic (original media with localized AI-generated edits). Video models classify into four: real, synthetic, semisynthetic, and rendered. Audio models classify into two: real and synthetic. Each modality is scored independently; miners may submit to multiple modalities by registering separate hotkeys.

Can I run GAS locally or on a private network?

GAS runs on Bittensor Subnet 34 (mainnet). Installation requires Python 3.10+, uv, Git, and system dependencies (ffmpeg, build tools, Node.js/PM2 for process management). Validators require an NVIDIA GPU with 80 GB VRAM (A100 recommended) for generation services. Discriminative miners only need to submit model files; generative miners need access to generation APIs or local GPU resources. Docker Compose is provided for validator setup with shared volumes for caches and wallet data.

What datasets does GAS use for benchmarking?

GAS evaluates against multiple sources: public benchmark datasets (image, video, audio) available in GASBench, private holdout datasets (not visible to miners), and continuously updated GAS-Station data from generative miners. Datasets are refreshed weekly. The round configuration assigns target weights to public, private, and GAS-Station samples. See GASBench's documentation for the full taxonomy and metric formulas.

BitMind (SN34) — Media Provenance Detection

Detecting synthetic media—deepfakes, AI-generated images, and manipulated audio—has become a critical challenge as generation technology evolves faster than detection tools. Static detection systems quickly become obsolete, and centralized approaches lack the scale and incentives to keep pace with adversarial improvements in generation quality. Organizations need detection classifiers that adapt as threats evolve, but building and maintaining these models in-house is expensive and isolated from the latest synthetic media research.

A decentralized network of detection contributors and synthetic-media generators creates a self-sustaining adversarial loop: detection models are continuously challenged by new synthetic content, forcing detectors to improve; better detectors push generators to create more realistic media; this cycle drives innovation in both domains without relying on a single vendor or research team. Participants compete for rewards based on the accuracy and robustness of their contributions, creating economic incentives aligned with the goal of staying ahead of synthesis threats.

GAS (Generative Adversarial Subnet) on Bittensor implements this model as an open competition. Discriminative miners submit detection classifiers for image, video, and audio that are evaluated against diverse real-world and synthetically generated datasets. Generative miners run servers that produce synthetic media on demand, earning rewards when their content passes validation and fools discriminators. Validators benchmark both types continuously, maintaining datasets updated with the latest synthetic examples and scoring models using calibration-aware metrics (MCC and Brier score) that reward both accuracy and confidence reliability.

Media forensics teams can submit detection models without hosting inference infrastructure, receiving continuous benchmark scores against public and private holdout datasets to validate model generalization.
Content moderation platforms can integrate top-ranked detection classifiers from the network to automatically flag synthetic media at scale, with scores updated weekly as new data arrives.
Synthetic media researchers can generate diverse test cases through the network's generative miners and immediately evaluate their detection improvements against real adversarial examples.
Security organizations can contribute proprietary detection expertise as models, earning rewards for accurate provenance classification while keeping model weights private (only predictions are evaluated).
Data labeling services can participate as generative miners, producing synthetic media with C2PA signatures that validators accept as training data for future detection model development.

GAS (Generative Adversarial Subnet) is a Bittensor subnet (SN34) where discriminative miners submit media-provenance classifiers for image, video, and audio detection, while generative miners create synthetic media to challenge those detectors. The adversarial loop drives innovation in both synthetic media generation and detection capabilities, with continuous benchmarking against diverse datasets including real-world and GAS-Station generated content.

The GAS subnet runs three parallel components: (1) Discriminative Miners submit detection models (safetensors format) evaluated by validators on diverse real and synthetic media across image/video/audio. Models scored via sn34_score and ranked via King of the Hill competition. (2) Generative Miners operate FastAPI servers responding to prompt requests from validators; generated content must pass C2PA signature verification and is scored on pass rate and adversarial fool rate. (3) Validators run three PM2/Docker services: the validator (sends challenges, scores both miner types, runs King of the Hill), the generator (creates synthetic media for discriminator evaluation using external APIs: OpenAI, OpenRouter, Stability AI, Runway), and the data service (manages benchmark datasets, handles model uploads). Discriminative miners do not host inference; generative miners must produce C2PA-signed content.

Emissions split 40% image, 40% video, 4% audio, 16% generators, with each discriminator lane awarded 85/10/5 to current/previous/prior-previous kings.

Media forensics teams can submit detection models without hosting inference infrastructure, receiving continuous benchmark scores against public and private holdout datasets to validate model generalization.
Content moderation platforms can integrate top-ranked detection classifiers from the network to automatically flag synthetic media at scale, with scores updated weekly as new data arrives.
Synthetic media researchers can generate diverse test cases through the network's generative miners and immediately evaluate their detection improvements against real adversarial examples.
Security organizations can contribute proprietary detection expertise as models, earning rewards for accurate provenance classification while keeping model weights private (only predictions are evaluated).
Data labeling services can participate as generative miners, producing synthetic media with C2PA signatures that validators accept as training data for future detection model development.

Emissions split 40% image, 40% video, 4% audio, 16% generators, with each discriminator lane awarded 85/10/5 to current/previous/prior-previous kings.

Discriminator sn34_score = √(M_norm · B_norm) where M_norm = clip((M+1)/2, 0, 1)^1.2 and B_norm = max(0, (B0-B)/B0)^1.8. MCC uses Gorodkin multiclass (or binary collapse). Brier measures calibration error.

Generator base reward R_base = p·min(n,10) where p=pass rate, n=verified samples. Fool multiplier M = max(0, min(2.0, f·s)) where f=fool rate and s=sample size bonus (ramps 0.5x at n<20, logarithmic to 2.0x at n≥20). Final reward R_total = R_base·M.

Model cost affects multiplier: sqrt(model_price/baseline_price). Discriminator emissions follow King of the Hill: 85% current king, 10% previous king, 5% king before that (per modality). Challenger needs ≥0.01 higher score on same benchmark version to dethrone.

BitMind

Registration Activity

BitMind

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Registration Activity

What does BitMind do?

Use Cases

About

Subnet History

Frequently Asked Questions

Development Activity

Recent Activity

Hardware Requirements

What does BitMind do?

Use Cases

About

Subnet History

Frequently Asked Questions

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Hardware Requirements

Miner Setup

Validator Setup

Overview

Registration Activity

Participate

Overview

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Registration Activity

What does BitMind do?

Use Cases

About

Subnet History

Frequently Asked Questions

Participate

Development Activity

Recent Activity

Hardware Requirements

What does BitMind do?

Use Cases

About

Subnet History

Frequently Asked Questions

Subnet Configuration

Market & Economics

For Miners

For Validators & Stakers

Hardware Requirements

Miner Setup

Validator Setup