Pbft

Resources

Byzantine General’s Problem

CS198.2x Week - Byzantine Fault Tolerance

Involved parties must agree on a single strategy to avoid complete a complete failure, but some of the involved parties may be corrupt or unreliable. The corrupt/unreliable nodes are called Byzantine nodes and there’s no solution to the problem when >= 1/3 nodes are byzantine.

  • Fault tolerant systems: Applicable when nodes can crash, not return values, but crash is detectable.
  • Byzantine tolerant systems: Applicable to fault tolerant systems and when nodes send incorrect/corrupted values.

Practical Byzantine Fault Tolerance (pBFT)

paper CS198.2x Week 1 - Practical Byzantine Fault Tolerance

PBFT handles < 1/3 byzantine faults/nodes. Has three phases: “pre-prepare”, “prepare”, and “commit”

  1. client sends information to primary node (e.g. derrick)
  2. node 2, nadir drops out due to network troubles
  3. Phase 1 - Pre-prepare: Derrick sends messages to all nodes (rustie, gloria, nadir)
  4. Phase 2 - Prepare: A node accepting a pre-prepare messages responds by sending a Prepare message to everyone else. A node is prepared when it has seen the pre-prepare message, sent its prepare message and has received 2f prepare messages from other nodes (leading to a total of 2f+1 prepares)
  5. Phase 3 - Commit: Nodes send out commit messages, if a node received f+1 valid commit messages they process the client requests then reply back to the client
  6. Client needs to wait for f+1 results

(f are adversarial nodes, but how do we know the # of f? maybe assume 1/3?)

Sybil Attacks

PBFT considers consensus but not sybil attacks. I.e. in the generals problem, N byzantine nodes could be the same corrupt general. Nakamoto Consensus handles sybil attacks via requiring block generation ability to be proportional to computational power available through the proof-of-work mechanism (i.e. it’s a unique consensus mechanism because it’s baked in, but PBFT and many other consensus mechanisms don’t have baked in Sybil resistance).

There are however straightforward solutions like using “weighted users”1

  1. See page 1 of Algorand paper 

Notes mentioning this note

There are no notes linking to this note.