https://github.com/BitMind-AI/bitmind-subnet

1. Introduction

The V2 release of Subnet 34 marks a major leap forward in decentralized deepfake detection. At its core is our most anticipated feature since our August mainnet debut: AI-generated video detection. This milestone release also incorporates weeks of careful architectural refinements.

This article serves as a comprehensive overview of v2.0.0, from its high level implications down to the nuts and bolts of its refactored codebase.

Let’s dive in!

1.1 V1.2.9 to V2.0.0

The integration of AI-generated video detection, the centerpiece of our V2 release, marks a crucial expansion of our subnet’s detection capabilities. In SN34 versions 1.x, all miner-hosted models operated on images, with no ability to capture the temporal patterns present in sequences of video frames. V2 closes this gap by providing video data transmission infrastructure, and allowing miners to deploy models that take temporal features into account during inference.

As a precursor to true video detection in BitMind’s applications, API calls to v1.x of SN34 would send the thumbnail frame of a video to the subnet miners for classification. While this anecdotally worked better than expected, V2 is positioned to dramatically improve our performance on synthetic video detection.

BitMind’s Deepfake Detection Applications

BitMind’s Deepfake Detection Applications

While our miners ramp up, competing to deploy the most accurate video detection algorithms, the BitMind team is working to bring video submission as a feature to our subnet’s collection of derivative applications. These two paths will soon converge to surface highly accurate, free, decentralized detection of AI generated videos to consumers.

1.2 Preamble

TLDR, Highlights of V2 include:

1. Expanded modality coverage to video
   1. VideoSynapse implementation for transmission of video challenges
   2. Three text-to-video models for synthetic video challenges, two datasets containing over 1 million real videos with significant semantic diversity for real dataset challenges
   3. Video base miner implementation, which includes a VideoDataset abstraction, dataset creation scripts, and detector training script
   4. Rewards implemented as a weighted combination of image and video performance, with an initial balance of 90% image, 10% video
2. Decoupled synthetic data generation from our validator process, which substantially simplifies the the validator’s challenge generation flow and improves generation throughput
3. Introduced a new image and video dataset caching system, which avoids increasing minimum validator storage to several terabytes

The next two sections are organized as follows:

• Section 2 provides an overview of the challenge landscape in V2, providing just enough detail to understand challenge generation and scoring.
• Section 3 walks through each component of the subnet’s architecture, some of which are also described in section 2, and provides details on how they have been refactored, optimized or extended in V2.