Owner: @Alexander Mozeika @Mehmet @Marcin Pawlowski

Attack scenario: Subnetwork control by the encoder/adversary

Architecture assumption

• There are N subnetworks, each responsible for storing one column.
• Each subnetwork contains multiple DA nodes, and (ideally) multiple independent nodes can serve the same column.
• First consider case of the same number of nodes and columns (Check Analysis of Impact of Omissions on Sampling Robustness , Screenshot 2025-08-01 at 08.58.41.png and Improving analysis of sampling robustness ).
• A consensus nodes performs DA sampling by requesting S columns (each request maps to the corresponding subnetwork).

Adversary goal and capabilities

Goal: influence DA outcomes by controlling the subnetworks involved in sampling and/or storage.

We distinguish capabilities:

1. Subnetwork capture (hard control).

   The adversary controls all DA nodes in some subset of subnetworks, so it can:

   • withhold responses (availability denial),
   • respond with incorrect data/proofs,
   • coordinate responses to bias the validator’s decision.

2. Partial presence (soft control).

   The adversary controls some DA nodes in many subnetworks, but honest DA nodes also exist and can answer.

3. Encoder equivocation / strategic behavior.

   The adversary behaves honestly during dispersal to get the blob included, then later attempts to reduce availability by dropping/changing data on controlled subnetworks, or by DoS-ing answers from those subnetworks.

This is the model I have in mind @Mehmet @Marcin Pawlowski. We have the number $n$ of DA nodes in total and out of these the number $n_A$ of nodes are controlled by adversary (red circles). The DA nodes are distributed (randomly) into $N$subnetworks (squares) in such way that each subnetwork has $R$ nodes. By chance all $R$ nodes of a subnetwork can be controlled by adversary (red square).