Neuron registration trends
The alpha price in τ, day by day, and what moved it each way.
Log change over the last 30 days −57.32%: protocol buys 0.00%, root sells −4.77%, users net −43.48%, cross-subnet −9.06%, unattributed −0.000000000000000000035%.
The chart follows the alpha price in τ. Each period starts at its opening price and steps through the mechanisms below, in a fixed order, to its closing price: a step that rises pushed the price up, one that falls pushed it down, and its height is how far it moved the price. A thin line carries each close into the next period’s open; it is dashed where the two differ. The faint candle behind each period is the market’s own open, high, low and close, for context.
With many periods on screen (90 days, or hours), the steps would be too thin to read, so the chart draws the price as a line instead and colours each period by the mechanism that moved it most. Hover a period for the full step-by-step breakdown.
Not here, because they do not move the spot price: the per-block TAO injection into the pool, and burns.
All figures are log changes, not simple percentage changes, so the parts add up exactly to the total; on big moves the two differ (a 10% log fall is a 9.5% price fall).
2026-11-09 UTC, as of ≈11:07 UTC (block 9,468,225)
The day isn’t over yet, so these figures keep changing. They are in no daily total, average or rank.
Organizations building computer vision systems face a critical gap: while image classifiers achieve high accuracy on clean data, they remain vulnerable to imperceptible perturbations—tiny pixel-level changes invisible to humans that cause misclassifications of medical scans, safety systems, and fraud detection models. Red teaming these vulnerabilities is expensive, manual, and static; most organizations lack the compute and expertise to systematically discover adversarial examples at scale.
A decentralized network of adversarial researchers can solve this continuously. When attack success is objectively verifiable (run the model, check the prediction, measure the perturbation norm), you can incentivize a competitive supply of attackers rather than maintaining a salaried red team. Each miner independently discovers stronger attacks; each validator cryptographically verifies the work in seconds; the network automatically rewards the best performers with on-chain emissions.
Perturb implements this on Bittensor by running a two-minute task cadence where miners fetch the current ImageNet-1k image, run adversarial attack algorithms, and submit perturbed images with pixel-level hashes. Validators download submissions, verify constraints (perturbation norm, image quality, prediction change) using EfficientNetV2-L, compute numerical scores, report results to a leaderboard API, and set on-chain weights from stake-weighted consensus across all validators' rankings.
Auto-generated from the GitHub repository and may not reflect the latest changes.
Perturb is a decentralized adversarial robustness network on Bittensor where miners compete to find imperceptible perturbations that cause image classifiers to fail, while validators construct challenges from real images, verify responses with mathematical precision, and reward the best attackers. The network produces two outputs: a continuously growing corpus of adversarial training examples and on-chain auditable proof of adversarial model evaluation.
The subnet has two main roles: miners and validators. Miners poll the task API for the current ImageNet-1k image, run local adversarial attack algorithms (baseline: PGD-style), upload perturbed images to configured storage, and submit image URLs with pixel-level hashes. Validators synchronously fetch the current task, wait for miner submissions via the API, download and verify each submission locally (checking constraints, image hashes, model predictions), compute numerical scores, report results to the leaderboard API, and periodically set on-chain weights based on stake-weighted consensus rankings from all validators' leaderboard reports.
The fixed reference model (EfficientNetV2-L) runs on every validator; miners can also fine-tune and commit models on-chain for the separate adversarial training track, evaluated once daily by validators against ImageNet-1k and adversarial datasets. Task generation is separated from validator runtime and publishes one task every 2 minutes; burn rate is fetched from a policy endpoint and assigned to a configurable burn UID at weight-setting time.
Per-response scoring applies hard gates (perturbation norm within [min_linf_delta, max_linf_delta], SSIM >= min_ssim, PSNR >= min_psnr_db, label change from original) then computes: linf_score = (1 - normalized_linf_ratio)^2, rmse_score = (1 - normalized_rmse_ratio)^2, perturbation_score = weighted_avg(linf_score, rmse_score) using PERTURB_LINF_COMPONENT_WEIGHT (0.7 default) and PERTURB_RMSE_COMPONENT_WEIGHT (0.3 default). Margin score and novelty score add 3% and 1% contributions respectively. Final score = PERTURB_PERTURBATION_WEIGHT * perturbation_score + margin_weight * margin_score + novelty_weight * novelty_score.
Failed verification yields 0.0 score. For model track, validators evaluate all committed models daily on ImageNet-1k validation and latest adversarial dataset; the best model receives the entire model share. Consensus model winner is determined by stake-weighted aggregation of validator evaluation reports.
The sections above were auto-generated from the GitHub repository and may not reflect the latest changes. View repository (github.com, opens in a new tab, external site)
Decentralized adversarial robustness network
This information was auto-extracted from the GitHub repository and may be incomplete or out of date. Always refer to the subnet's GitHub repository (github.com, opens in a new tab, external site) for the latest instructions.
Not yet verified: Notes
Not yet verified: GPU, GPU model, VRAM
Python 3.10+, NVIDIA driver and CUDA stack compatible with PyTorch required
Machine-extracted from the subnet's repository docs; may be wrong or out of date. Check the repository before buying or renting hardware. Open the repository (github.com, opens in a new tab, external site) · Extracted 5 Oct 2026
Clone the Perturb repository
git clone https://github.com/0xsigurd/Perturb && cd Perturb
Install PM2, create Python virtual environment, and install dependencies for miner role
bash ./scripts/setup_common.sh miner
Create miner environment configuration file
cp scripts/miner.env.example scripts/miner.env
Configure required miner runtime parameters
Edit scripts/miner.env with required fields: WALLET_NAME, WALLET_HOTKEY, NETUID, NETWORK
Start miner using PM2
bash ./scripts/run_miner.sh
Clone the Perturb repository
git clone https://github.com/0xsigurd/Perturb && cd Perturb
Install PM2, create Python virtual environment, and install dependencies for validator role
bash ./scripts/setup_common.sh validator
Create validator environment configuration file
cp scripts/validator.env.example scripts/validator.env
Configure required validator runtime parameters including Hugging Face token for ImageNet-1k access
Edit scripts/validator.env with required fields: WALLET_NAME, WALLET_HOTKEY, PERTURB_API_KEY, HF_TOKEN
Verify HuggingFace token can access the gated ImageNet-1k dataset
python scripts/check_imagenet1k.py
Start validator using PM2
bash ./scripts/run_validator.sh
| Day (UTC) | Share |
|---|---|
| 2026-11-08 | 0.00% |
| 2026-11-07 | 0.00% |
| 2026-11-06 | 0.00% |
| 2026-11-05 | 0.00% |
| 2026-11-04 | 0.00% |
| 2026-11-03 | 0.00% |
| 2026-11-02 | 0.00% |
| 2026-11-01 | 0.00% |
| 2026-10-31 | 0.00% |
| 2026-10-30 | 0.00% |
| 2026-10-29 | 0.00% |
| 2026-10-28 | 0.00% |
| 2026-10-27 | 0.00% |
| 2026-10-26 | 0.00% |
| 2026-10-25 | 0.00% |
| 2026-10-24 | 0.00% |
| 2026-10-23 | 0.00% |
| 2026-10-22 | 0.001% |
| 2026-10-21 | 0.0025% |
| 2026-10-20 | 0.01% |
| 2026-10-19 | 0.01% |
| 2026-10-18 | 0.01% |
| 2026-10-17 | 0.01% |
| 2026-10-16 | 0.01% |
| 2026-10-15 | 0.01% |
| 2026-10-14 | 0.01% |
| 2026-10-13 | 0.01% |
| 2026-10-12 | 0.01% |
| 2026-10-11 | 0.01% |
| 2026-10-10 | 0.01% |
| 2026-10-09 | 0.01% |
| 2026-10-08 | 0.01% |
| 2026-10-07 | 0.01% |
| 2026-10-06 | 0.01% |
| 2026-10-05 | 0.01% |
| 2026-10-04 | 0.0044% |
| 2026-10-03 | 0.00% |
| 2026-10-02 | 0.00% |
| 2026-10-01 | 0.00% |
| 2026-09-30 | 0.00% |
| 2026-09-29 | 0.00% |
| 2026-09-28 | 0.00% |
| 2026-09-27 | 0.00% |
| 2026-09-26 | 0.00% |
| 2026-09-25 | 0.00% |
| 2026-09-24 | 0.00% |
| 2026-09-23 | 0.00% |
| 2026-09-22 | 0.00% |
| 2026-09-21 | 0.03% |
| 2026-09-20 | 0.03% |
| 2026-09-19 | 0.03% |
| 2026-09-18 | 0.03% |
| 2026-09-17 | 0.03% |
| 2026-09-16 | 0.04% |
| 2026-09-15 | 0.03% |
| 2026-09-14 | 0.04% |
| 2026-09-13 | 0.03% |
| 2026-09-12 | 0.04% |
| 2026-09-11 | 0.03% |
| 2026-09-10 | 0.03% |
| 2026-09-09 | 0.04% |
| 2026-09-08 | 0.05% |
| 2026-09-07 | 0.04% |
| 2026-09-06 | 0.05% |
| 2026-09-05 | 0.06% |
| 2026-09-04 | 0.06% |
| 2026-09-03 | 0.06% |
| 2026-09-02 | 0.05% |
| 2026-09-01 | 0.05% |
| 2026-08-31 | 0.05% |
| 2026-08-30 | 0.05% |
| 2026-08-29 | 0.05% |
| 2026-08-28 | 0.05% |
| 2026-08-27 | 0.06% |
| 2026-08-26 | 0.06% |
| 2026-08-25 | 0.06% |
| 2026-08-24 | 0.06% |
| 2026-08-23 | 0.05% |
| 2026-08-22 | 0.06% |
| 2026-08-21 | 0.08% |
| 2026-08-20 | 0.07% |
| 2026-08-19 | 0.08% |
| 2026-08-18 | 0.07% |
| 2026-08-17 | 0.07% |
| 2026-08-16 | 0.07% |
| 2026-08-15 | 0.08% |
| 2026-08-14 | 0.05% |
| 2026-08-13 | 0.01% |
| 2026-08-12 | 0.01% |
| 2026-08-11 | 0.0016% |
