Is Stellar As Secure As You Think?

Yujin Potter, Minjeong Kim, Yujin Kwon, Yongdae KimEuros&p Workshops 2019 · DOI

TL;DR

This paper investigates the security of Stellar's consensus protocol (SCP), a variant of Federated Byzantine Agreement (FBA). The authors demonstrate that Stellar's quorum slice structure is highly centralized, making it vulnerable to cascading failures. Specifically, they show that the failure of just two nodes, controlled by the Stellar Foundation, can bring down the entire network.

What Problem Does This Paper Address?

Stellar is a popular cryptocurrency that uses a consensus mechanism called Federated Byzantine Agreement (FBA). FBA is designed to be more flexible than traditional Byzantine Fault Tolerance (BFT) algorithms, particularly for open-membership blockchains where nodes can join and leave freely. A key aspect of FBA is the "quorum slice," a set of nodes that a given node trusts for consensus. The security of FBA, and thus Stellar, heavily relies on the structure of these quorum slices.

While FBA offers flexibility, its security is highly dependent on how nodes configure their quorum slices. The authors identify a critical gap: there's been no comprehensive study on the actual structure of quorum slices in the live Stellar system and how this structure impacts its security. This paper aims to fill that gap by analyzing Stellar's quorum slice structure, measuring node influence, and evaluating the system's resilience to failures, particularly cascading failures.

Methods

The authors employ a multi-pronged approach:

  • FBA Analysis: They first theoretically analyze FBA, proving that it is not inherently superior to traditional BFT algorithms like PBFT in terms of safety and liveness. They introduce the concept of an (f, x)-FT system to quantify tolerance to faulty nodes, showing that FBA's x-value (maximum fraction of faulty nodes tolerated) can range from 0 to 33%, which is not better than PBFT's theoretical maximum of 100%.
  • Data Collection and Analysis: They collect data on Stellar's quorum slices from various sources (e.g., mystellar.tools, stellarbeat.io) over several months. This data includes validator IP addresses, quorum slice configurations, and lists of existing validators.
  • Node Influence Measurement: To quantify the influence of individual nodes within the quorum slice structure, they use two metrics:
  • PageRank (PR): A standard algorithm for ranking nodes in a graph, adapted from web page ranking.
  • NodeRank (NR): A newly proposed metric that considers not only the number of incoming edges (trust) but also the influence of the nodes providing that trust and the threshold values of the quorum slices.
  • Cascading Failure Simulation: They simulate cascading failures by systematically removing nodes from the network and observing the impact on consensus. This involves identifying pairs of nodes whose failure leads to widespread consensus failure.

Datasets

The paper uses data collected from the live Stellar network. Specifically, they gathered information on:

  • Validator IP addresses
  • Generated quorum slices
  • Lists of existing validators

This data was collected over a period leading up to January 29, 2019, and also from public websites like mystellar.tools, stellarbeat.io, and stellar.org. They also note that 18 nodes were already offline at the time of their analysis.

Key Insights and Ideas

  • FBA is not inherently superior to PBFT: The theoretical analysis shows that FBA's flexibility doesn't translate to better worst-case security guarantees than traditional BFT.
  • Stellar's quorum slice structure is highly centralized: This is the paper's central finding. Despite Stellar's open-membership design, the actual network configuration is dominated by a few influential nodes.
  • The Stellar Foundation nodes are disproportionately influential: The analysis reveals that nodes run by the Stellar Foundation (e.g., sdf_validator1, sdf_validator2, sdf_validator3) have the highest NodeRank and PageRank scores, indicating significant control over consensus.
  • Small quorum slices exacerbate centralization: A large percentage of quorum slices contain ten or fewer validators, making them more susceptible to failure.
  • Trust model leads to centralization: The paper argues that Stellar's reliance on a trust model, where users manually select trusted validators, naturally leads to a preference for popular or affiliated nodes, thus reinforcing centralization.
  • Cascading failures are a severe threat: The centralized structure makes Stellar highly vulnerable to cascading failures, where the failure of a few nodes triggers a chain reaction leading to network-wide consensus failure.

Key Results

  • Centralization confirmed: Data analysis shows that 58% of quorum slices have ten or fewer validators, and 83.9% of validators are directly related to Stellar (including the Foundation and business partners).
  • Node influence bias: The sdf_validator nodes consistently ranked highest in both PageRank and NodeRank, with similar NodeRank values, suggesting similar influence.
  • Catastrophic cascading failure: The failure of just two nodes, specifically any pair of sdf_validator nodes, causes 100% of the network to fail to reach consensus. Even the failure of eno and sdf_validator1 leads to over 90% failure.
  • Stellar is a (2, ~4.5)-FT system: Based on their findings, the authors estimate Stellar's tolerance to faulty nodes to be approximately 2 nodes causing a ~4.5% failure rate (considering active validators), which is significantly worse than PBFT.

Limitations and Future Work

  • Limitations:
    • The analysis was conducted on data from January 2019, and the network structure may have evolved since then.
  • The paper acknowledges that while FBA is not superior to PBFT in theory, a well-structured FBA system could be secure. The issue lies with Stellar's current structure.
  • The authors note that attracting major financial institutions to participate as validators is challenging, which could improve decentralization.
  • Users rarely change their quorum slices, making dynamic adaptation to failures (like in PBFT) impractical for Stellar.
  • Future Work:
    • The paper suggests that to improve decentralization, the number of validators needs to increase, and they should be included in slices more evenly. However, they note that increasing validators is difficult due to the lack of internal incentivization.
  • Mitigation strategies like lowering quorum slice thresholds are discussed, but they come with trade-offs (reduced safety).
  • The authors imply that a more robust incentivization mechanism or a shift in user behavior regarding quorum slice management would be necessary for Stellar to achieve a more decentralized and secure state.

Summary updated

BibTeX
@inproceedings{potter2019is,
  author = {Yujin Potter and Minjeong Kim and Yujin Kwon and Yongdae Kim},
  title = {{Is Stellar As Secure As You Think?}},
  booktitle = {Euros\&p Workshops},
  year = {2019},
  doi = {10.1109/eurospw.2019.00048},
}
Figure