Research
Attacking P2P Networks: Liveness Attacks on the GossipSub and DOG Gossip Protocols
Shows constant-size latency-advantaged adversaries can split honest nodes in GossipSub and DOG gossip by gaming peer scoring, making Ethereum's gossip liveness a concrete protocol-risk item.

- arXiv
- 2611.03509
- Published
- 2026-11-02
- Authors
- Jiaheng Li, Anton Paramonov, Yann Vonlanthen
AI summary
Overview
Research area: Peer-to-peer network security, specifically liveness attacks against blockchain gossip protocols (GossipSub, used by Ethereum, and DOG, designed for CometBFT). Technical level: Advanced — the paper relies on spatial latency models, peer-scoring mechanics, and probabilistic analysis, though the attack intuition is describable in plain terms. Scope: The paper analyzes how an adversary who never forges or alters any message can still starve honest nodes of fresh messages by exploiting the peer- and route-selection mechanisms of two gossip protocols.
What This Paper Is About
Blockchains depend on a gossip layer to move transactions, blocks, and consensus messages between nodes, but the security of that layer is studied far less rigorously than the consensus protocols sitting on top of it. The authors ask whether an adversary whose only goal is to break liveness — preventing honest nodes from receiving new messages — can win by manipulating the protocol's own peer scoring and route filtering rather than by attacking cryptography. They answer yes for both GossipSub and DOG, using small amounts of attacker infrastructure and only legitimate protocol actions.
Key Contributions
- A latency-based takeover of GossipSub peer scoring. The paper shows that an adversary with a constant number of geographically distributed nodes and a fixed multiplicative latency advantage can dominate the First Message Deliveries scoring term and thereby hold every score-selected mesh slot of a target node.
- A formal spatial model with separation conditions. Nodes are placed in the unit square with latency modeled as L₂ (Euclidean) distance, and the authors derive control-rate conditions α_u > T and per-link conditions β_{u,p} > T with T = D_score/(D_score+1), proving the resulting separated fraction is independent of network size n.
- Treatment of GossipSub's countermeasures. The paper analyzes the random-peer selection mechanism, the outbound-connections quota (including a connectivity argument showing that preserving outbound connections alone keeps the honest graph connected, so the quota must be attacked separately), and clustered node distributions.
- A single-node attack on DOG. The paper gives counterexamples where DOG's route filtering disconnects the dissemination graph with no failures and no Byzantine participants, then presents a practical attack using one physical node with a constant number of identities (five logical nodes suffice under the default configuration) that disables every forwarding route across the whole network.
Main Findings
-
GossipSub separation is feasible with a constant attacker footprint. Under the analyzed geographical model, a constant number of geographically distributed attacker nodes plus a fixed latency advantage (q) separates a constant fraction of nodes from the ordinary mesh, before accounting for the outbound-connections quota.
-
Defaults used throughout are Prysm's. The reference configuration is D = 8, D_lo = 6, D_hi = 12, D_score = 4, D_out = 2.
-
The threshold has two forms. The basic condition uses T = D_score/(D_score+1) = 4/5 (ε = 0), and a more conservative choice uses ε = 0.5, giving T = 6/7.
-
Numerical scale. Simulations sample n = 1000 honest nodes and 100 attacker nodes uniformly from [0,1]², sweeping k ∈ {1, …, 100} and q ∈ {0, 0.5, 0.8}.
-
~20% full separation at the conservative threshold. With q = 0.8 and k = 100, approximately 20% of nodes satisfy α_u > 6/7.
-
Single-link removal is far easier. For the same parameters, approximately 98%–99% of ordered pairs satisfy β_{u,p} > 6/7, meaning most individual honest links can be pruned even though full node separation is rarer.
-
Iteration amplifies the attack dramatically. Modeling messages from already-separated nodes as automatically attacker-controlled, 100 independent runs with n = 1000, k = 100, q = 0.8 yield an initial separated fraction of roughly 40% at T = 4/5, rising to 100% separation in all 100 runs; at T = 6/7 roughly half the runs reach 100%.
-
Random peer slots can be ground down. Up to D_hi − D_score peers may sit outside the score-protected attacker set, but repeated grafting past D_hi forces re-pruning and resampling, which attacker identities can win over many repetitions given enough distinct IP addresses to avoid the IP-colocation penalty.
-
The outbound quota blocks the basic attack on its own. A union-bound argument shows that an outbound-only graph with d ≥ 2 outbound peers per node stays connected with high probability as n → ∞, so inbound-only grafting does not achieve separation.
-
Attacking the quota requires presence in the replacement pool. The adversary must hold a small fraction of identities in the candidate pool; the worked example uses 100 geographically distributed attacker nodes, q = 0.8, and attacker identities at 1%–2% of the pool, assuming uniform sampling of the pool.
-
Clustered nodes are easier, not harder. Deploying one attacker node per data-center cluster gives very low latency to that cluster's honest nodes, producing a stronger delivery advantage than for uniformly distributed nodes.
-
DOG collapses completely but not instantly. The attack disables all forwarding routes and can be maintained indefinitely, yet direct publisher-to-neighbor delivery over mesh links still works — a new message reaches the publisher's direct neighbors but cannot be forwarded beyond them.
-
DOG attack costs. Establishment cost scales with the number of two-hop routes (source, one intermediate forwarder, target); maintenance cost scales with the number of fresh messages published, plus forwarding work proportional to each publisher's direct neighbor count.
-
Unlimited publication is a convenience, not a requirement, for DOG. The principal construction grants unlimited fresh messages, but the attack can proceed gradually at limited publication frequency or reuse honest-originated messages in some stages.
Methodology in Plain English
The authors build a deliberately simplified map of the world: honest nodes are scattered uniformly at random over a unit square, and the time for a message to travel between two nodes is approximated by the straight-line distance between them. Attackers are placed on the same map but get a discount factor q on communication among themselves, representing paid-for better infrastructure. A message is "controlled" by an attacker if the attacker can physically deliver it to the target faster than the honest peer holding the target's mesh slot — including the hop through the attacker's own private network. Counting what fraction of all possible message sources an attacker can beat in this race gives a control rate per target, and if that rate clears a threshold derived from the number of score-selected mesh slots, the attacker's peers survive pruning while honest peers are eventually dropped. The authors then check these rates numerically by simulation, and they add an iterative step where nodes already captured become free message sources for capturing further nodes. For DOG, the approach is different: they construct explicit counterexamples to route filtering and then build and simulate a concrete attack, reporting cost, assumptions, and mitigations.
Why This Matters
Consensus safety proofs assume messages eventually arrive. This paper shows that assumption can be voided by an adversary who never breaks a signature and never needs a majority of stake or hash power, which shifts the security budget of a blockchain to include the gossip layer explicitly.
Real-world applications:
- Ethereum attestation and block propagation. Attackers exploiting the First Message Deliveries term could isolate validators from the mesh, delaying inclusion and skewing nodes' views.
- CometBFT-based chains using DOG. Disabling all forwarding routes turns a global broadcast network into isolated publisher-to-neighbor islands, which can stall transaction dissemination.
- Staking and validator operations. Operators need to know that a modest, geographically distributed adversary footprint (100 nodes in the analyzed example) can be enough to erode their connectivity.
- Client teams and testnets. The conditions and thresholds give client implementers concrete parameters (D_score, D_out, decay settings) to stress-test against.
Industry relevance: the attacks require only ordinary participation, rented infrastructure, and optionally a better network link, making them cheap relative to their impact — exactly the kind of low-cost, high-leverage threat that protocol maintainers must parameterize around.
Future Directions
- The IHAVE/IWANT metadata layer is explicitly left open. Separation from the mesh does not prove a node receives nothing via metadata exchange, and the mechanisms forming that network are specified largely outside GossipSub, so they are not analyzed.
- Proving the iterative process's fixed point. The authors show empirically that iteration can reach 100% separation for some parameter choices and sampled configurations, but state that a general analysis of the final fixed point is beyond the report's scope.
- Hardening the outbound-connections quota. The quota argument depends on uniform peer sampling and a candidate pool containing all nodes or a uniform sample, plus attacker presence at the 1%–2% level; designing replacement sampling that resists this is an obvious follow-up.
- Mitigations for DOG. The paper promises analysis of possible mitigations but the truncated content does not report their details, leaving concrete defenses and their evaluation as open work.
Target Audience
Protocol researchers and client developers working on Ethereum, CometBFT, and other gossip-based systems; security engineers assessing peer-to-peer attack surface; and graduate students in distributed systems or cryptography and security who want a worked example of how liveness can fail without any cryptographic break. Readers need comfort with graph models and scoring mechanics, though the qualitative conclusions are accessible to a technically literate non-specialist.
Authors’ abstract
Peer-to-peer (P2P) networks form the communication layer of large decentralized systems such as Ethereum, and the correctness of higher-level protocols depends on the gossip layer delivering messages to honest nodes. We study the robustness of two gossip protocols against adversaries whose only goal is to violate liveness, that is, to prevent honest nodes from receiving fresh messages, without forging, altering, or invalidating any message. For GossipSub, the…
The full abstract on arXiv