Scribd
Upload
201 views

Symbolic MEV Extraction

The document discusses using symbolic execution and property-based testing to extract MEV from Ethereum smart contracts without white-box knowledge of their logic. It proposes defining properties like balances not decreasing, symbolically executing transactions to generate constraints, and using an SMT solver like Z3 to find inputs that break the properties, potentially enabling MEV extraction. Future work could include developing Rust-based symbolic EVM interpreters and EVM-specific SMT solvers, or using reinforcement learning to improve solver heuristics for faster MEV searches.

Uploaded by

isnull
Copyright
© © All Rights Reserved
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
201 views

Symbolic MEV Extraction

The document discusses using symbolic execution and property-based testing to extract MEV from Ethereum smart contracts without white-box knowledge of their logic. It proposes defining properties like balances not decreasing, symbolically executing transactions to generate constraints, and using an SMT solver like Z3 to find inputs that break the properties, potentially enabling MEV extraction. Future work could include developing Rust-based symbolic EVM interpreters and EVM-specific SMT solvers, or using reinforcement learning to improve solver heuristics for faster MEV searches.

Uploaded by

isnull
Copyright
© © All Rights Reserved
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 20

Symbolic MEV Extraction

[email protected]
@gakonst
Disclaimer: early brainstorm, I’m not a searcher,
might be impractical, open-ended discussion
Question: Can we apply techniques used for software security
in MEV extraction?
State of MEV Searching (for the most part)
● Main categories:
○ Atomic / stat / top of block arbitrage
○ Backrunning / sandwiching
○ Liquidations
○ “Long tail”
● Software Stack:
● Geodistributed nodes
○ Fast access to state (e.g. Uniswap reserve values)
○ Strong mempool access (Eg. peer directly with AVAX validators)
● Handwritten bot for each integration
○ “White box” simulations (E.g. native Uniswap in Rust, vs going to the EVM)
● Reliable submission & closing (e.g. PGA / Flashbots auction)
● Observability
State of MEV Searching (for the most part)
● Main categories:
○ Atomic / stat / top of block arbitrage
○ Backrunning / sandwiching
○ Liquidations
○ “Long tail”
● Software Stack:
● Geodistributed nodes
○ Fast access to state (e.g. Uniswap reserve values)
○ Strong mempool access (Eg. peer directly with AVAX validators)
● Handwritten bot for each integration
○ “White box” simulations (E.g. native Uniswap in Rust, vs going to the EVM)
● Reliable submission & closing (e.g. PGA / Flashbots auction)
● Observability
Black vs Grey vs White Box MEV Extraction
● Black Box: No knowledge of the insides of a system
● Grey Box: Some knowledge
● White box: Full knowledge

In MEV extraction:
● White box: custom code per app, maintenance as apps change, reliable
● Black box = property-based, unclear how effective bc of no knowledge of the
app
● Grey box = property-based, extracts info about the app as it runs
But Ethereum is a Dark Forest
● Not all bots are application specific
● “Generalized” frontrunners
○ Simulate others’ transactions
○ Look at state diff + balance changes
○ If any subcall is profitable → copy + frontrun
● Pros:
○ Can cover a wider range of opportunities / long tail mev discovered by others
● Cons:
○ Unclear how effective it really is vs specialized
○ Relies on frontrunning / copying, does not discover new opportunities
Can we build a bot that reliably extracts MEV on arbitrary contracts
without knowledge of application logic?
Property-based Testing

Testing kinds:
● Unit testing: test pre-defined scenarios
● Property-based testing: check that certain conditions hold over multiple inputs

Two ways to use:


● Defensively: Ensuring your code is working as expected
● Aggressively: Find counterargument that breaks a property
Property-based MEV Extraction

1. Define properties for contract solvency: e.g. ETH/token balance does not go
down
2. Monitor for new contract bytecodes
3. For each contract
a. Get all its functions from the jumptable
b. Call them with various arguments many times
c. Find arguments which break the defined properties
4. Submit series of transactions required to break the property
5. ???
6. $$$
Property-based MEV Extraction

1. Define properties for contract solvency: e.g. ETH/token balance does not go
down
2. Monitor for new contract bytecodes
3. For each contract
a. Get all its functions from the jumptable
b. Call them with various arguments many times
c. Find arguments which break the defined properties
4. Submit series of transactions required to break the property
5. ???
6. $$$
What inputs? Random!?
Symbolic Execution

1. Translate your “concrete” problem into a problem of constraints


2. SMT solvers can find solutions to problems with constraints better than you
can by randomly throwing info at something
Symbolic Execution
● Concrete: x = 2, y = 3 → execute
● Fuzzing/Property-based: assign random
x,y → execute
● Symbolic:
○ Track symbolic state via SMT queries
■ x, y, z = 2 * y
○ Track constraints
■ 2 * y == x
■ x <= y + 10
○ Send to solver (z3 / ccv3 / …)
○ Solver outputs TRUE or FALSE + input
which breaks
Symbolic EVM Execution
● Explore all possible states of an EVM execution
● Used for formally verifying smart contracts
○ AND output the inputs when properties break
○ E.g. during audits
● Examples
○ HEVM
○ Pakala
○ Mythril
○ Manticore
○ ..?
● Maybe we can use this to make $?
https://fv.ethereum.org/2020/07/28/symbolic-hevm-release/
https://www.palkeo.com/en/projets/ethereum/pakala.html
https://manticore.readthedocs.io/en/latest/evm.html
Symbolic MEV Extraction
● Monitor new contracts
● Symbolically execute multiple EVM transactions on that contract → generate
constraints
● Send the constraints to Z3
● ??
● $$$

Black box, but much better than random because the constraints identify
the structure of the problem, which z3 can solve effectively.
..people can get creative
..people can get creative
The Concrete vs Symbolic Searcher
“The Concrete” THE SYMBOLIC
Hasn’t checked the
trends in weeks
Looks for alpha

Knows every
Rewrites solidity contract inside outThe alpha finds him
contracts in rust for
speed
Only does arbs, Doesn’t know Solidity Writes his properties
liqs & top of once
block
Maintenance
takes all his time Writes his own
Code is 5 years old Lets the solver do his job
bots
Future: Fast Symbolic MEV Extraction
● Rust Symbolic EVMs
○ https://github.com/WilfredTA/symbolic-stack-machines (goal to integrate in Foundry)
○ https://github.com/williamberman/evm-symbolic-execution/blob/master/EVM%20Symbolic%20
Execution.ipynb
● EVM-specific SMT solvers
○ https://github.com/EVM-SMT/solver
● Solvers are largely heuristic based → teach & generalize the heuristics via
reinforcement learning? (e.g. train an RNN vs z3)
● Hardware acceleration: MEV ASICs?! (really SMT ASICs)
○ Probably a bad idea, z3 & others are heuristics based so if they take a long time they will
never find the answer, h/w will not help
Thank you for your attention!
Q&A?

[email protected]
@gakonst

You might also like