Thin Client Security

From Bitcoin Wiki
Revision as of 16:16, 3 December 2012 by Eldentyrell (talk | contribs) (Depth and Difficulty)
Jump to: navigation, search

Recently there have been a number of proposals for bitcoin clients which do not store a complete copy of every block in the entire block chain. This page will refer to all such clients as "thin clients". This page is meant to be a place to try to make sense of the security and trust implications of the various schemes.

Block Height vs. Depth

It is important to distinguish between block height verification and block depth verification.

A client verifies the height H of a block by checking that there are H block before it, all of which are well-formed and obey the maximum-difficulty-adjustment-rate rule. Currently only the Satoshi client and libbitcoin do block height verification. Block height is the fundamental anchor of trustless security in the Bitcoin system.

A client verifies the depth D of a block by checking that there are D blocks after it (also called "confirmations"), all of which are well-formed. SPV clients substitute block depth for block height as a transaction validity check. All clients use block depth as a measure of the liklihood of a blockchain reorganization producing a new longer fork which excludes the transaction (i.e. orphaning its block).

See also some comments on probabilistic verification of block height.

Full-Chain Clients

The "thick" bitcoin client downloads a copy of the entire chain, including all transactions (not just headers). It will be used as the reference point for security comparisions below.

Block Height as a Transaction Validity Check

A full-chain client trusts the difficultywise-longest well-formed blockchain it can find. Any transaction on the difficultywise-longest well-formed chain is considered valid. Therefore, the validity of a transaction is determined by its height -- i.e. how many blocks come before it. A transaction's depth (the number of blocks after it) is used to determine the likelihood of the transaction being invalidated due to the emergence of a longer fork.

Header-Only Clients

These client downloads a complete copy of the headers for all blocks in the entire blockchain. This means that the download and storage requirements scale linearly with the amount of time since bitcoin was invented; it would be preferable to have the scaling be logarithmic or even constant.

Simplified Payment Verification (SPV)

This scheme is described in section 8 of the original bitcoin whitepaper.

Block Depth as a Transaction Validity Check

As Satoshi writes, "[the thin client] can't check the transaction for himself, but by linking it to a place in the chain, he can see that a network node has accepted it, and blocks added after it further confirm the network has accepted it." If we take "X" to be the "number of blocks added after it", then SPV essentially trusts that a transaction X blocks deep in the chain does not have inputs which were already spent further back in the chain. Therefore, the validity of a transaction is determined by its depth -- i.e. how many blocks come after it. Other thin client protocols also include this assumption.

This is very different from the trust model in the "thick" client: the thick client verifies that a transaction's inputs are unspent by actually checking the whole chain up to that point -- there is no "X blocks deep" involved here. The thick client uses "X blocks deep" (aka "confirmations") only once it has already decided that a transaction is valid (i.e. no double-spends). At that point it uses "X blocks deep" to decide how likely it is that a longer fork in the chain will emerge which excludes that transaction.

It is very important to understand how the same property ("X blocks deep") is used to verify two different properties in the thick client and SPV cases. The thick client never uses block depth as a measure of transaction validity; the SPV client does.

This is a concern in a situation where an SPV client is subjected to a double-spend attack by somebody who controls its network connection. For example, suppose you are at a wi-fi cafe and are paying for something using your smartphone -- the cafe owner controls your network connection. Satoshi acknowledges this implicitly when he writes that "the verification is reliable as long as honest nodes control the network" -- to be completely pedantic, this means that the verification is reliable as long as honest nodes control the part of the network that the SPV client is able to communicate with. In an attack-by-ISP scenario this may not be a sufficiently strong security property. The attacker would not need to overpower "the rest of the network" because the client is unable to communicate with it.

Depth and Difficulty

The main problem with using block depth as a validity check is difficulty.

It is trivial for to forge a transaction which is 100 blocks deep if each of those blocks can be difficulty=1.0 blocks. A thin client might poll nearby nodes to find out what the difficulty ought to be (accepting the last hundred or so blocks as proof), but this approach is acutely vulnerable to Sybil attacks. Therefore, clients that use block depth as a validity check need to have some sort of hardwired assumption about the difficulty level. It's easy for a human to say "well, the difficulty ought to be plausible," but formalizing this in software is not as easy as it seems.

  • The difficulty levels of late 2012 seem implausibly high from the perspective of the CPU-mining era of late 2009.
  • Future chip technologies could increase the network hashrate by factors of 10 or more.
  • Although SHA-256 is unlikely to be totally compromised, it is entirely possible that a future mathematical discovery could shave a dozen or so bits off of the hashspace, corresponding to a 100x increase in hashpower. Similar partial-collision attacks have been discovered for many earlier hash functions.
  • A sudden drop in exchange rate might lead to a mass exodous of miners, causing the network hashrate to plunge.
  • Some future competing cryptocurrency might become more popular, leading to a sudden plunge in the bitcoin network hashrate.

All of these factors and others mean that any sort of difficulty-guessing heuristic wired into a piece of software will make it fragile. This is an especially serious concern for embedded software or software operated by bitcoin non-experts. Imagine if a bitcoin ATM or vending machine suddenly stopped working because of an unexpected change in the network hashrate! It is unlikely that the person servicing this sort of equipment would be able to figure out what had gone wrong.

Clients which use height rather than depth (by full chain validation or probabalistic height verification) are not vulnerable to this problem. By tracking the chain continuously as it develops they learn about difficulty adjustments from evidence which is astronomically expensive to forge, even as technological and economic factors alter the exact definition of "expensive".


Simplified Payment Verification is the verification mechanism used in BitCoinJ.

A security analysis of some of the issues in BitcoinJ can be found here; however:

  • The claim that "picking 10 nodes and requiring all of them to be consistent needs much less trust" overlooks the problem of "cancer nodes" and Sybil attacks.
  • Many of the security claims are qualified by some form of "if you don't think an attacker controls your internet connection"; see the previous section for a discussion of why this is problematic.


Simplified Payment Verification is the verification mechanism used in picocoin.

The library (libccoin) that picocoin is based on includes code for validating scripts and blocks; this could potentially be used to implement a full-chain client.

Unused Output Tree in the Blockchain (UOT)

There have been several proposals (the first appears to be this one by gmaxwell, who called it an "open transaction tree", although the term "open" is now taken to mean "not yet mined into the blockchain" rather than "unspent") to form a tree of unused transaction outputs at each block in the chain, hash it as a Merkle tree, and encode the root hash in the block chain (probably as part of the coinbase input). This will be called an Unused Output Tree (UOT). The first detailed proposal so far appears to be Alberto Torres' proposal; etotheipi's ultimate blockchain compression is a variant of this.

If such UOT hashes were included in the blockchain, a client which shipped with a checkpoint block that had a UOT would only need to download blocks after the checkpoint. Moreover, once the client had downloaded those blocks and confirmed their UOTs, it could discard all but the most recent block containing a UOT.

This would also let a thin client reduce the question of "is this output unspent" to the question of "is this block super-well-formed" where "well-formed" means "well-formed according to the normal blockchain rules and additionally has an Unused Output Tree which is accurate and truthful". This is still a long way from the low level of trust involved in the thick client, but it is a major improvement over all existing proposals.

It is unlikely that bitcoin would ever arrive at a state where every single block had a UOT, since this would require upgrading 100% of the miners on the network, or else convincing enough miners to reject blocks which do not contain a UOT. The latter strategy risks creating blockchain forks, which can be expensive (in reward terms) to miners. Therefore, any UOT strategy would need to cope with the fact that not every block contains a UOT.

Hostile miners may insert blocks into the chain which have what claims to be a UOT, but which is actually invalid. It is unlikely that such blocks could be kept out of the chain because, again, this would require adding a new block well-formedness criterion, and miners implementing this new criterion would risk "mining on the wrong side" of a fork, which could cost them a lot of money. Therefore, any UOT strategy would need to cope with the fact that not every block containing a UOT entry can be trusted.

Note that at the present moment no standard format for such Unused Output Tree hashes has been agreed upon, nor do any of the blocks in the chain contain them. The ultraprune feature added to bitcoind-0.8 maintains a similar data structure on the client's disk. It does not put this data structure or its hash anywhere in the blockchain.

Server-Trusting Clients

These clients involve some (usually low) level of trust in the server they rely upon. Mechanisms for authenticating the server, and for confirming that the server has not been compromised, are usually not explained.

All thin clients listed below currently connect to a single server, and are vulnerable to an attack similar to a double-spend. The attack can be run by that single server - the server can just lie to them that they received a Bitcoin transaction, and they, assuming the server does not lie, perform some service, transfer funds or send goods without actually receiving any Bitcoin in exchange. Therefore, they are implicitly trusting it.

Future enhancements have been suggested that will have the client talk to multiple servers and broadcast transactions and query all of them. Unfortunately it is well known to security researchers that this does not actually increase security; it simply makes the exploits more complicated and difficult to find. Security researchers have a name for this phenomenon: it is called a "Sybil attack"[1]. This post on bitcointalk explains how some governments (notably Iran and China) already perform these sorts of attacks on their own citizens, with the coerced assistance of SSL certificate authorities.

Clients with a checkpoint (even a very old one) that download and validate the headers for the whole blockchain are not vulnerable to Sybil attacks in the following sense: they can always ensure that an attack would cost more than the amount being stolen.