Security Analysis of Consensus Protocols in Adversarial Settings

Abstract

Consensus protocol security proofs tell you what an adversary cannot do in theory. They say nothing about what a clever adversary can actually achieve in practice -- degrading throughput, inflating latency, censoring transactions, and manipulating ordering -- without ever violating the safety property that the proof guarantees. This gap between theoretical safety and practical resilience is where real attacks live. We present the Adversarial Consensus Security Framework (ACSF), a systematic security evaluation of six consensus protocols -- PBFT, HotStuff, Tendermint, Avalanche, Bullshark, and Raft -- under nine adversarial attack strategies that exploit practical weaknesses without breaking theoretical safety bounds. Our attacks include strategic leader delay, selective transaction censorship, MEV-style ordering manipulation, bandwidth exhaustion, eclipse attacks on subsampled voting, and adaptive corruption. We introduce the Adversarial Resilience Score (ARS) measuring throughput degradation, latency inflation, censorship resistance, ordering fairness, and recovery speed under each attack. Key finding: DAG-based protocols (Bullshark ARS 0.924) are structurally more resilient than leader-based protocols because the DAG architecture eliminates the single leader bottleneck that most practical attacks target. Tendermint shows the best resilience among leader-based protocols (ARS 0.812) due to its round-robin rotation and gossip-based dissemination.