On the Incentive Compatibility of Bitcoin & Cryptocurrency Loi Luu Joint works with Jason Teutsch, Raghav Kulkarni, Ratul Saha, Inian Parameshwaran, Aquinas Hobor & Prateek Saxena National University of Singapore Bitcoin is becoming more important Total market: 4 Billion USD Number of Bitcoin research papers More investment 250 – Venture Capital Funding for Bitcoin Startups 205 Triples in 2014 200 – Growing 25% faster than the internet in its early 150years More adoptions 100 – Paypal, Microsoft, Dell 61 – Bank of Lodon 50 – Nasdaq and MAS interested in21Blockchain 8 More academic research 0 1 0 1 2008 2009 2010triples 2011 in 2014 2012 – Research in Bitcoin 2013 2014 2 Contents Bitcoin’s background Incentive-compatibility in cryptocurrency protocol (CCS’ 15) Incentive-compatibility in Bitcoin pooled mining protocol (CSF’ 15) 3 BITCOIN 101 4 Ideal Bank Account Functionality Bank “Send $2 from my account to Bob.” Alice Alice: Bob: $08 $10 -2 $22 $20 +2 Ledger “You’ve got Money! $2 from Alice.” Bob Ideal Bank properties • Alice cannot spend money that she doesn’t have • Bank cannot send the money without Alice’s acknowledgement • Bank cannot keep the money without sending to Bob • Bob should be able to spend the money Slides from Andrew Miller From Ideal Bank to Bitcoin in 5 Steps 1. Implement the Bank as a trusted third party (e.g., Paypal) Alice Bank Bob 2. Implement the Bank as a multiparty computation P1 Alice P2 P5 P3 P4 Slides from Andrew Miller Bob - Standard results in Byzantine faulttolerance apply here, (e.g. Paxos) - PKI is assumed From Ideal Bank to Bitcoin in 5 Steps 3. Suppose we have a magic token that chooses parties at random. ? Alice ? ? ? Bob ? Whoever has the token gets to broadcast *once* • If t parties are malicious: Pr[honest selected] = (n-t)/t • Thm. If majority are honest, transaction log converges Slides from Andrew Miller *caveats From Ideal Bank to Bitcoin in 5 Steps 4. Replace the token with computationally hard Puzzle - Solvable by concurrent/independent participants - No advantage over brute force Scratchd(puz, m): r ← {0,1}k; if H(puz || m || r) < 2k-d then return r ? Alice Slides from Andrew Miller ? ? ? ? Bob From Ideal Bank to Bitcoin in 5 Steps 5. Finally, provide participation incentives • give each “lottery winner” a reward • also solves the problem of initial allocation • Incentive compatible participation? ? Alice Slides from Andrew Miller ? ? ? ? Bob Slightly More Detail • Ledger: state file, mapping amounts of BTC to pkeys • Transactions: Signed instructions to modify the ledger • Blockchain: Authenticated sequential log of transactions Each solution is used as seed for the next puzzle challenge. The solutions form linked lists (blockchains). Thm. For all n, eventually converge on unique n-length chain. Slides from Andrew Miller Bitcoin system overview Miners (Validate TXs & generate blocks) TXs TXs Users (generate TXs) Blockchain Mining Bitcoins in 5 easy steps 1. Join the network, listen for transactions a. Validate all proposed transactions 2. Listen for new blocks, maintain blockchain a. When a new block is proposed, validate it 3. Assemble a new valid block 4. Find the nonce to make your block valid a. SHA256(BlkTemplate || Nonce) has D leading zero bits, e.g.: 0000000000000000024f37840… 5. When find a valid block a. Broadcast & hope it gets accepted b. Receive reward Bitcoin transaction Specify the source of the money Prove of eligibility to spend Input: PreviousTX: ID of previous transaction Index: 0 scriptSig: Sign(PubKey), PubKey Output: Amount to send Who to send to and what payee has to do to spend Logic of the transaction Value: 5000000000 scriptPubKey: %take Signature and PubKey as params checkif Hash(PubKey) = Payee's ID, checkif Sign(PubKey) is valid Bitcoin script: supports limited operators • Prevent DoS attack • Easy to verify • Limit the applications Ethereum: Cryptocurrency with Turing-complete script • Can run arbitrary program on blockchain – Enable more applications • Introduce Smart Contract (SC) – A public program that embeds contractual clauses between parties – Has its own address, local storage, etc. – User triggers SC by sending a transaction if msg.datasize==2: return msg.data[0] + msg.data[1] if msg.datasize==1: if SHA256(msg.data[0]) == contract.storage[1]: send(reward, msg.sender) Ethereum system overview TXs TXs TXs Smart Contract INCENTIVE-COMPATIBILITY IN CRYPTOCURRENCY PROTOCOL Incentive in Bitcoin protocol Incentive for miners – Block reward – Transaction fees included in the block There is no reward for block verifier! – “When a new block is proposed, validate it” People verify other’s block because – They want to mine valid blocks – For the “common good” – Normally, its cheap 18 Steps to verify a block If block hash meets difficulty – One SHA256 computation Merkle tree of TXs is correctly constructed – O(No.OfTXs) SHA256 computations If all TXs are valid – Depends on number of TXs – Logic each TXblock: Currently in in a Bitcoin - N=500-700 TXs - Verifying a normal TX requires 1 signature, 1 SHA256 - Thus, verifying a Merkle tree is cheap What would happen if verifying a block were not cheap? 19 Problem Is cryptocurrency protocol incentivecompatible? – Incentivize miners to verify block? – Are honest miners vulnerable? Finding: Cryptocurrency protocol is not incentive compatible – Miners are vulnerable to resource exhaustion attack – Rational miners have incentive to skip verifying block 20 Contribution Establish that cryptocurrency protocol is not incentive compatible – Verifier’s dilemma Formalize the cryptocurrency consensus protocol – Understand the incentive structure Propose an incentive compatible solution – Techniques to deploy proposed solution in existing cryptocurrency – Case studies: Outsourced computation applications 21 Resource exhaustion (RE) attack Attacker creates block that requires long time & much resource to verify – Bitcoin: Block that has many TXs – Ethereum: TX that has infinite loop Damage – Attacker gets higher chance in finding next blocks – DoS attack other miners Existing mitigations – Bitcoin: Limit block size ~ 1 MB • Limit no. of TXs – Ethereum • Gas fee charged as the amount of opcodes executed – Make REA expensive for attacker • Gas_limit to limit block execution Is this enough to prevent the attack? 22 RE attack in Bitcoin Intuition: Bitcoin limits the blocksize, but not the number of opcodes – Expensive opcode ~ easy opcode • SHA256, CheckSig, etc – What if a TX requires 10000 signatures verification? The attack: CVE-2013-2292 – Attacker includes multiple OP_Checksig in a block-size TX – Miners have to hash 19.1 GB to verify • Take relatively 190 seconds CPU-time • Expected time to find a block is only 10 mins 23 RE attack in Ethereum Intuition – The gas fee is credited to the block founder • Attacker = block founder? – gas_limit can be adjusted by miners The attack – Creates expensive smart contract SC – Sends a TX to activate SC N = matrix_size – Include TX in his own block A = N*N input matrix – Others have to run SC when B = N*N input matrix if msg.data[0] = 1: verifying his block C = get_matrix(msg.data[1]) – Attacker conducts the attack if (C == A * B) //run O(N3) sendReward() with 0-fee 24 Verifier’s dilemma Miners do not know whether to verify a block – Verify and be vulnerable to RE attack – Not verify and mine on top of invalid blocks TXs and computations may be incorrect Miners also have incentive to skip block verification – Gain advantage in the next race – Avoid RE attack Existing cryptocurrency protocols are not incentive compatible 25 The problem is real and immediate - 5% miners mine an invalid block - ~Half the network hash rate was mining without fully validating blocks - Build new blocks on top of that invalid block. 26 Our solution CRYPTOCURRENCY AS A CONSENSUS VERIFIABILITY PROTOCOL 27 Consensus verifiability model A consensus verifiability (CV) protocol – – – – G: Problem giver asks a solution for f(x) P: Prover proves that he has a solution s V: Verifier verifies if s=f(x) is correct Wblk: work that V always does to get reward Bitcoin as a CV – G: sender decides what receiver has to do to spend – P: receiver proves the ownership of the address – V: verify if receiver’s signature is valid CV in Ethereum – G can define more expressive problem f() – V may have to do more work 28 Threat model: ε- rational miner Def 1: Advantage of rational miner adv(f) = Wf - Wdf – Wf: amount of work that verifying f() requires – Wdf: amount of work in deviated protocol – Generally adv(f) = Wf – O(1) Def 2 Advantage to skip block verification adv(blk) = å adv( fi ) - O(1) = i£N åW fi - O(1) i£N Def 3:ε- rational miners are honest if • adv(blk) ≤εWblk • deviate otherwise 29 Incentivize correct consensus verifiability Def 4:ε- consensus verifiability is a CV that requires at most εWblk in verifying a block Lemma 1:ε- consensus verifiability is incentive compatible w.r.t ε- rational miners εvalue • Represents the acceptable “common good” work • Not straightforward to estimate, depends on • Net-worth of applications • The network properties • The incentive mechanism • Individual miner’s beliefs 30 Achieve ε-CV in existing cryptocurrencies Goal: limitingεWblk work in verifying a block eWblk Method: Limiting work in each TX to wtx = N0 – In Ethereum • Leveraging the gas function G(W) – Determine the upper bound on the gas required to do W work • Only allows TXs requiring less than G(wtx ) gas – In Bitcoin • Introduce TX size • Bound number of expensive opcodes • Only allow standard TXs How about applications that require more than εWblk work computation? 31 Porting more applications to ε-CV: Correct consensus verifiability Split verification work into smaller TXs – Each TX fits in ε-CV model – Advantage of rational miners is bounded – Correctness guaranteed – Latency may be high N = matrix_size A = N*N input matrix B = N*N input matrix Each TX will if msg.data[0] = 1: check only C = get_matrix(msg.data[1]) one element if msg.data[0] > 1: i, j = get_index(msg.data) check_if (C[i][j] == A[i][] * B[][j]) //require to run O(N) 32 Porting more applications toε-CV: Approximate consensus verifiability Sacrifice correctness to achieve low latency with probabilistic checking – reduce number of samples, thus TXs and latency – can only guarantee correctness to a certain extent Intuition – if a solution y’ is deemed correct y’ ~ f(x) Goal – Ensure y’ differs from f(x) by at most δbits with at least prob. of p (say, 99%) • At mostδbits in y’ have different property required in f(x) with prob. ≥p • y’ is computed from x with prob. ≥p 33 Case studies: Outsourced computation Correct consensus verifiability – GCD computation of large numbers – Dot product Approximate consensus verifiability – Matrix multiplication – Sorting – k-coloring 34 Conclusion Bitcoin and existing cryptocurrencies are not incentive-compatible – Verifier’s dilemma – Consensus computation may be done incorrectly Formalize the consensus protocol – Understand the incentive structure – Propose incentive compatible solutions Techniques to deploy large applications in the proposed solutions – Achieve correctness – Achieve performance 35 INCENTIVE-COMPATIBILITY IN POOLED MINING Pooled mining Mining: Requires huge computational power – Hardware investment: >100 millions USD – Miners have to wait for years! Delegation of computational power via pooled mining – Pooled supervisor distributes work and reward – Miners find share • Find Nonce to have d (<D) leading zeros – Eg: 000000123fa… • Shares are meaningful to pool only More than 90% are pool miners – Pool miners get frequent reward 0011X 0010X 0001X 0000X Securing Bitcoin pool protocol is important! 37 Problem Is Bitcoin pooled mining protocol secure? – Miner’s reward computational power? – Following the protocol best outcome? Intuitive answer: Yes – Hash inversion is cryptographically hard This work – Shows an attack to make a million USD per month 38 Block Withholding Attack ● A topic of hot debate – ● Even from a pool operator – ● “Withholding attacks don’t make financial sense — that’s easy to prove with math...” “Basically in no way has an accurate model of the network shown withholding to be more profitable than legitimate mining...” Still happen in practice – The attack caused a damage of 200, 000 USD to Eligius pool Our findings - The attack does profit the attacker - Applicable to all cryptocurrencies 39 Contributions Study the Bitcoin pooled mining protocol – Game theoretic approach, i.e. formulate Bitcoin mining as a game Analyze the BWH attack – The attack is profitable • Pool protocol is vulnerable – Empirically evaluate the findings 40 Model BITCOIN MINING AS A COMPUTATIONAL POWER SPLITTING GAME 41 Compete to get 25 BTCs D=4 d=2 Find 0000X 25 BTCs Free to distribute power 5 BTCs 42 Bitcoin as a Computational Power Splitting Game N pools Player: α • Player action: Pick =(β0, β1, β2 ,…, βn) – Use αβ0 to compete independently – Contribute αβi to pool Pi – Get reward Ui from pool i • Player’s goal is to maximize R = åUi GAME NETWORK P1 P2 αβ1 αβ0 αβ2 … Pn-1 Pn αβi PLAYER αβn 43 Case study BLOCK WITHHOLDING ATTACK 44 Block Withholding Attack ● ● Only submit “normal” shares – Reduces pool’s reward and other miners’ reward – Pool has to pay the attacker for his shares Hard to detect – Finding a block is probabilistic Honest BWH 0011X 0011Y 0010X 0010Y 0001X 0001Y 0000X 0000Y 45 BWH attack is profitable Intuition: Bitcoin is a zero-sum game – Coins supply is constant – The loss in the victim pool is picked up by other pools BWH attack +x + X -x -0.2X +0.8X 46 Simple example attacker BWH attack 5% Victim pool 75% 20% 25% 75% Attack Scenario Honest Scenario 5% 1 pool, α=25% (β0, β1) = (0.8, 0.2) αβ0 = 20% αβ1 = 5% 0% 75% 20% 79% 21% Honest Scenario Mining Power Actual Mining Power Distribution Reward Honest Attack scenario scenario Attacker 25% 25% 25.9% Pool 75% 75% 74.1% 4.9% 21% 74.1% Actual Reward Distribution 47 Analyze BWH attack using CPS game Compute the reward of the attacker – Before vs after the attack in each pool – Infer attacking rules Consider different scenarios – Single attacker, single pool – Single attacker, multiple pools – Multiple attackers 48 Scenario: single attacker Extra reward ab (p - b1 ) DR = p(1- ab1 ) 1 Victim pool’s size Attacking portion Attacker’s power It’s always profitable to BWH attack $b1 < p : D R > 0 There is a threshold on the attacking power D R > 0 Û b1 < p It’s more profitable to target big pool Exists the optimal strategy to maximize D R 49 Other scenarios There are other dishonest miners – It’s possibly profitable – Depends on how much the pool is “contaminated” Attacking multiple pools – Attacks as many as possible – Exists the optimal strategy 50 Nash equilibrium What is the best strategy for the miner? Consider two accessible pools – The dominant strategy is to attack the other There is no pure strategy – There is always a better move to win back BWH from P1 BWH from P2 P1 P2 51 Does attack’s duration matters? Does it actually profit? 1111BTCs/ BTCs/1210mins mins 10 BTCs/ 10 mins • Short term • It depends • Long term • Yes • Difficulty adjusts 52 Evaluate our results ● ● Use “official” Bitcoin client, popular pool mining software – Run on cloud-based Amazon EC2 – Burning up to 70,000 CPU core-hours Essential to – check the correctness of our result – show our CPS model is faithful 53 Experimental results Attacker’s Power Attack Scenario Rewar d 25% One pool 25.66% 30% One pool 31.14% 45% One pool 46.9% 25% Multiple pools 26.49% Relative difference: 1% 100% 90% 0 0 0 0.1 0 0 0 0 0.09 0.08 0.15 0.14 0 Unknown 28% 0.20 70% 60% 0 0.1 0.09 0.08 50% 40% 26% 0.08 0.07 0.13 24% 0 0.2 30% 0.18 0.17 0.15 0.14 A acker Reward Propo on of mining power 80% 30% KNC Ghash Discus Fish A acker 0.13 20% 22% Reward 1 10% 0.6 0.64 0.67 0.62 0.64 0.53 0% 20% Before a ack 1st change 2nd 3rd 4th 5th change change change change 6th change 54 Discussion on Defenses Assign same task to multiple miners Change pay-off scheme – pay more to shares which are valid blocks Change Bitcoin protocol to support pooled mining natively – Make share become oblivious to miner • only pool supervisor knows which shares are valid blocks A cheap and compatible solution to prevent BWH attack is still an open problem 55 Conclusion Security of pool protocols is an open research topic Existing pool protocols are vulnerable to BWH attack – Game-based model to understand incentive structure Future work – Defenses – Proof of security 56 Thank you Q&A Email: loiluu@comp.nus.edu.sg BTC LTC 57 Related work BWH attack – [Rosen11] Analysis of bitcoin pooled mining reward systems • Attack is not profitable – [CoBa14] On subversive miner strategies and block withholding attack in bitcoin digital currency • Attack does profit, but analysis is incorrect – [Eyal15] The miner’s dilemma • Arrives at same findings, but from pool perspective • No experimental evaluation • Concurrent work Other Bitcoin attacks – [Rosen11] • Pool hopping, Lie in wait attack – [EyalSi13] Majority is not enough: Bitcoin mining is vulnerable • Selfish mining attack 58 59