– Clients updates – liveness of DOG and GossipSub protocols in an adversarial scenario by Jiaheng Li
P2P Networking #007
Transcript
- Csaba
Spanning tree is most efficient under a few conditions: Static network conditions Abundant bandwidth
- Csaba
When you say latency advantage, is it to all nodes? Because that would be irrealistic.
- Yann Vonlanthen
Replying to "When you say latency..." No, just between the adversarial nodes.
- Yann Vonlanthen
Reacted to "Spanning tree is most efficient under a few conditions: Static network conditions Abundant bandwidth" with 👍
- Csaba
I see, its an adversarial fast-forward to gain advantage
- Yann Vonlanthen
Replying to "When you say latency..." Right, @Anton Nashatyrev also suggested that It could also be achieved simply by not checking signatures
- Anton Nashatyrev
Replying to "When you say latency..." Yep, no validation at all
- Csaba
Replying to "When you say latency..." Adversaries can always trust each other :-)
- Anton Nashatyrev
Reacted to "Adversaries can always trust each other :-)" with 😃
- Csaba
Replying to "When you say latency..." Also no need to respect fairness rules on the network, etc.
- kasey
Replying to "When you say latency..." They can trust not only their adversarial peers, but also can wager on random peers with a good track record.
- Anton Nashatyrev
Replying to "When you say latency..." @Csaba Do you mean like fairnes of TCP, QUIC congestion control algos?
- Csaba
Replying to "When you say latency..." Exactly, fairness on the underlay (but also, they not need to use the same underlay at all, can sit on the two sides a direct fibre).
- Yann Vonlanthen
Replying to "When you say latency..." Yes exactly 🙂
- Raúl Kripalani
Replying to "When you say latency..." i'd boil this down to: you can be wire protocol-compliant, yet have a very different behaviour than "official" clients
- Raúl Kripalani
Replying to "When you say latency..." and that's usually the area where behavioural exploits lie
- Raúl Kripalani
does this model take into account existing churn in the network, which in itself can "heal" pathological cases?
- Csaba
Replying to "does this model take..." That’s why we do “induced churn” in Geth ;-)
- Yann Vonlanthen
Replying to "does this model take..." Good point, but I think churn actually might help the attacker here, because you could get connected to an already eclipsed node, which would make things worse for you
- Csaba
Replying to "does this model take..." No, because you would move on (in the EL case, maybe not here)
- kasey
You should also see the message id for your unaggregated attestations long before they are included.
- Yann Vonlanthen
Reacted to "You should also see the message id for your unaggregated attestations long before they are included." with 👍
- kasey
(When you are subscribed to the subnet)
- Csaba
Reacted to "and that's usually the area where behavioural exploits lie" with 👍
Detailed notes
Client Updates
- •Prysm: Gloas partial column work dev-complete; interop tested with Lighthouse; rebasing against other Gloas work this week00:03:10
- •Prysm: full peer scoring revamp underway; design doc targeting next P2P call00:03:23
- •Lighthouse: considering enabling partial columns on mainnet in next release00:05:12
- •Geth: sparse blob pool optimizations — erasure coding speedup 50ms→4ms; serving fixed for custodied columns beyond first 64; protected peer slots now on master00:06:22
- •Geth: GoKZG performance PRs in progress (not yet in production); block diffusion design exploration underway (segmentation, erasure coding, push/pull modes)00:10:25
- •Geth: security issues from AI audit being documented; peer-forced cell/blob recomputation via repeated GetBlobs identified as attack vector00:12:24
- •Nethermind: IPv6 support improvements; sparse blob pool merging to master soon; boot node release targeting next Nethermind release00:13:40
Research Gossipsub Dog Liveness
- •DOG protocol: adversary with one physical node can disable all forwarding routes network-wide; even without adversary, counterexample shows connectivity not guaranteed00:14:57
- •GossipSub: under geometric latency model, ~100 adversary nodes with 80% latency advantage can separate ~40% of nodes; 99% of links vulnerable to targeted removal00:18:14
- •GossipSub outbound quota: adversary with 1% of identity pool can eventually take over both outbound mesh slots; iterative attack boosts separation rate further00:28:54
- •Latency advantage achievable by skipping signature validation or using direct fibre/low-latency network — wire protocol-compliant but behaviourally distinct00:36:18
- •Proposed mitigations: statistical delivery-latency feedback loops; self-health evaluation; attestation on-chain landing rate as outbound mesh quality signal; 'secret buddy' baseline comparison node00:43:49
Action items
- •Share Prysm peer scoring design doc with P2P working group before next callPrysm (ashah)
- •Share fast block diffusion design write-up with interested parties on requestGeth (Csaba)
- •Publish document summarising AI-audit security issues including peer-forced blob recomputationGeth (Bosul)
Agenda
From the call issue #2203ShowHide
Key decisions
AI Disclaimer: Some content or metadata on EIPsInsight may be AI-inferred or automatically compiled. If you find any discrepancy, please contact us at dev@avarch.org.