Release 3.1.0

[dylanuys]

Overview

The BitMind subnet has been expanded to support segmentation challenges in addition to the existing classification challenges. This update introduces a new miner type called SEGMENTER that specializes in identifying AI-generated regions within images through pixel-level segmentation.

Key Changes

1. New Miner Type: SEGMENTER

The subnet now supports two distinct miner types:

• DETECTOR: Performs classification tasks (real vs synthetic vs semi-synthetic)
• SEGMENTER: Performs segmentation tasks (identifying AI-generated regions)

2. Miner Type Declaration

Miners must declare their type through the /miner_info endpoint:

{
    "miner_type": "SEGMENTER",
    "uid": 123,
    "hotkey": "5FHneW46xGXgs5mUiveU4sbTyGBzmstUspZC92UhjJM694ty"
}

Important: Validators will query this endpoint to determine which type of challenges to send to each miner.

Segmentation Challenge Details

Task Description

SEGMENTER miners receive images that contain AI-generated regions (typically semi-synthetic images with AI modifications like faceswaps, inpainting, etc.). Your task is to:

1. Identify AI-generated regions: Detect which pixels in the image were created or modified by AI
2. Generate a confidence mask: Output a continuous mask where each pixel value represents your confidence that the pixel is AI-generated

Input Format

• Content-Type: image/jpeg or application/octet-stream
• Image dimensions: Variable (will be resized to 256x256 by validators)
• Color space: RGB

Expected Response Format

SEGMENTER miners must respond with:

1. Binary mask data: Raw bytes of a numpy array with shape (H, W) where H and W match the input image dimensions
2. Required headers:
   • X-Mask-Shape: Comma-separated height and width (can be static "256,256" for now.)
   • X-Mask-Dtype: Data type of the mask (e.g., "float16")
   • Content-Type: "application/octet-stream"

Response Structure

# Example response structure
heatmap = np.array([[0.1, 0.8, 0.2], ...])  # Shape: (H, W)
heatmap_bytes = heatmap.astype(np.float16).tobytes()

headers = {
    "X-Mask-Shape": f"{heatmap.shape[0]},{heatmap.shape[1]}",
    "X-Mask-Dtype": str(heatmap.dtype),
    "Content-Type": "application/octet-stream"
}

return Response(
    content=heatmap_bytes,
    headers=headers,
    status_code=200
)

Mask Requirements

• Value range: All values must be in [0.0, 1.0]
• Data type: float16 (recommended for efficiency)
• Shape: Must match the input image dimensions exactly
• Semantics:
  • 0.0 = High confidence the pixel is real/original
  • 1.0 = High confidence the pixel is AI-generated
  • Values in between represent uncertainty

Scoring Mechanism

Intersection over Union (IoU)

SEGMENTER miners are scored using the Intersection over Union (IoU) metric:

IoU = |Prediction ∩ Ground Truth| / |Prediction ∪ Ground Truth|

Where:

• Prediction: Your binary mask (thresholded at 0.5)
• Ground Truth: The true AI-generated regions provided by validators

Reward Calculation

1. Instantaneous IoU: Calculated for each challenge
2. Moving Average: Your reward is the average IoU over your recent challenges (up to 200 most recent)
3. Score Range: Rewards range from 0.0 (poor performance) to 1.0 (perfect segmentation)

To start, rewards will be split 80% to detectors, 20% to segmenters. This is both based on the amount of traffic detectors need to service to uphold the core functionality of our applications, as well as to take a progressive rollout approach to segmenters. This split is subject to change, and input from the community is welcome.

Performance Tracking

• Validators track your IoU scores over time
• Scores are stored per modality (currently only images)
• Historical performance affects your weight distribution

Implementation Guide

1. Environment Configuration

Update your .env.miner file to specify the miner type:

MINER_TYPE=segmenter  # or "detector" for classification

2. Running a SEGMENTER Miner

# The start_miner.sh script automatically detects MINER_TYPE
./start_miner.sh

3. Model Implementation

The segmentation miner template is located at neurons/miner/segmentation_miner.py. Key components:

Segmenter Class (neurons/miner/segmenter.py)

class Segmenter:
    def __init__(self, config):
        # Initialize your segmentation model here
        pass
    
    def segment(self, image_tensor):
        # Implement your segmentation logic
        # Return: np.ndarray of shape (H, W) with values in [0, 1]
        pass

Required Methods

If you use the default segmenter.py, you can implement these methods.

• preprocess(): Prepare input for your model
• segment(): Perform segmentation and return confidence mask
• postprocess(): Convert model output to proper format

You can also write your own class from scratch to be used in segmentation_miner.py.

Challenge Types

Current Support

• Modality: Images only (video segmentation not yet supported)
• Media Type: Semi-synthetic images (AI-modified content)
• Augmentation: Random augmentations applied to test robustness and mitigate gaming

Migration from Classification

If you're currently running a DETECTOR miner:

No immediate action required: Your existing miner will continue to work. Miners without the miner_info endpoint will be considered detectors that haven't added this endpoint yet. However, we would appreciate you adding this endpoint as soon as you're able.

Technical Requirements

Support and Resources

• Discord: Reach out to the team on Discord for support, to report bugs, or to suggest features.
• Documentation: Check the main Mining Guide for general setup

Important Notes

1. Registration: Registering on mainnet burns TAO - test thoroughly first
2. Performance: The provided template model is for demonstration only
3. Competition: Implement your own high-performance segmentation model
4. Scoring: Only image segmentation is currently supported
5. Updates: Monitor for future updates that may add video segmentation

---

Happy Mining! 🚀

For questions and support, reach out to the BitMind community on Discord.