Skip to content
Crypto.world

Research

PEB Separation and State Migration: Unmasking the New Frontiers of DeFi AML Evasion

Industrial crypto AML (Elliptic, Chainalysis, TRM Labs) and most academic work treat token transfers as a graph and assume that economically meaningful value movement can be reconstructed by tracing…

PEB Separation and State Migration: Unmasking the New Frontiers of DeFi AML Evasion
arXiv
2603.26290
Published
2026-03-27
Authors
Yixin Cao, Xianfeng Cheng, Yijie Liu

AI summary

Overview

Research area: Blockchain forensics and anti-money laundering (AML) for decentralised finance, filed under Cryptography and Security (cs.CR); arXiv:2603.26290v1, 27 March 2026, by Yixin Cao and Xianfeng Cheng (EigenPhi) and Yijie Liu (Independent Researcher).

Technical level: Intermediate. The paper mixes a concrete Ethereum case study and a hands-on mainnet-fork simulation with lightweight formal notation (transfer graphs, two definitions, two propositions and one informal theorem). Readers comfortable with DeFi primitives such as AMMs, limit orders and flash loans will follow it easily; the formalism is not heavy.

Scope (one sentence): The paper formalises two structural mechanisms in composable DeFi — Principal–Execution–Beneficiary (PEB) separation and state-mediated value migration — and argues that they make transfer-graph-based AML neither attribution-complete nor causally closed, illustrating each with a real transaction and a constructed cross-pool arbitrage validated on a mainnet fork.

What This Paper Is About

Industrial crypto AML (Elliptic, Chainalysis, TRM Labs) and most academic work treat token transfers as a graph and assume that economically meaningful value movement can be reconstructed by tracing paths through that graph. The authors argue this assumption collapses in composable smart-contract ecosystems, where the party who expresses an intent, the party who submits the transaction, and the party who receives the value can all be different, and where value can be moved purely by contract state transitions such as AMM reserve rebalancing without any direct transfer between the parties. The goal is to prove that this is a structural limit of the transfer-layer abstraction rather than a quirk of one protocol, and to argue for a shift toward AML based on execution semantics.

Key Contributions

  1. Formalises PEB separation for AML. The paper defines the functional decoupling of intent originators (Principals), transaction executors such as MEV searchers (Executors), and ultimate recipients (Beneficiaries), and shows how this breaks the link between risk-labelled actors and on-chain actions.
  2. Introduces state-mediated value migration. It describes an execution-level mechanism in which economic coupling is enforced through invariant-driven contract state transitions (for example constant-product AMM reserve updates) rather than through transfer continuity.
  3. Shows the two mechanisms jointly imply transfer-layer incompleteness. Through Proposition 1 (Role-Mediated Indeterminacy), Proposition 2 (State-Mediated Non-Closure) and an informal Theorem 4.1 (Transfer-Layer Incompleteness), the authors argue that even full trace visibility does not remove attribution ambiguity.
  4. Demonstrates structural generativity under composability. The paper argues that under modular composition and open, incentive-driven participation these failure modes are not corner cases but reproducible by construction, characterising a formal observational boundary for heuristic flow tracing.

Main Findings

  • Three implicit assumptions underpin transfer-based AML: flow continuity (value migration is isomorphic to the transfer layer), initiator visibility (suspicious conversions are initiated by high-risk addresses) and endpoint linkage (the recipient of a transfer is the ultimate beneficiary). The paper's central claim is that all three fail under composable execution.
  • A real transaction exhibits full PEB separation. In Ethereum Mainnet transaction 0x04c43669c930a82f9f6fb31757c722e2c9cb4305eaa16baafce378aa1c09e98e (Block 15,937,667), approximately 38.2M USDC were converted into 38.1M DAI via a 1inch limit order filled by an independent MEV bot, involving the 1inch Limit Order Protocol, Uniswap V3 and Aave flash liquidity. The Principal (a phishing-labelled maker), the Executor (the filler bot) and the Beneficiary (the DAI recipient) are functionally disjoint, written as P ∩ E ∩ B = ∅. The Principal never appears as msg.sender, and no direct token transfer occurs between maker and receiver.
  • Role-mediated indeterminacy (Proposition 1). Because the transaction is third-party-initiated, transfer-layer observation may fail to attribute beneficiary realisation to the principal even when economically enforced migration clearly occurred.
  • State-mediated causal non-closure (Proposition 2). When value relocation is enforced through execution-level state transitions that do not create a uniquely attributable directed path in the transfer graph, the transfer-layer abstraction is not causally closed with respect to economic migration.
  • Transfer-layer incompleteness (Theorem 4.1, stated informally; referred to as Theorem 1 in Section 5.5). In composable environments with modular execution and open participation, there exist execution patterns for which transfer-layer observations are neither attribution-complete nor causally closed. The authors stress this requires no protocol malfunction — only economically rational execution in standard DeFi architectures.
  • A constructive four-step arbitrage relocates value with no traceable path. Using two independent constant-product pools P₁ and P₂ over the same (A, B) pair, an Operator executes two back-to-back cross-pool loops: Phase 1 (TX1) injects (a + x) of asset A into P₁ for b of B, then b of B into P₂ to recover x of A and repay flash liquidity, with a sourced from the Principal via allowance; Phase 2 (TX2) extracts b′ of B from P₂ and returns b′ to P₁ for y + a′ of A, with y chosen so that a′ = a. The net effect is Δbalance_A(P) = −a and Δbalance_A(B) = +a with no direct transfer between P and B. Appendix A derives the exact consistency condition for the first loop: r_B^(1)(x+a)(r_A^(2) − x) = r_B^(2) x (r_A^(1) + x + a).
  • The construction works under realistic fees. An on-chain simulation on an Anvil fork of Ethereum Mainnet at Block 21808947 used Uniswap V2 WETH/USDT pool 0x0d4a11d5EEaaC28EC3F61d100daF4d40471f1852 and SushiSwap WETH/USDT pool 0x06da0fd433C1A5d7a4faa01111c044910A184553, both with a 0.3% swap fee, to launder a = 10.0000 WETH from EOA P to EOA B using a flash loan of x = 50.4893 WETH (written as 50.49 WETH in the setup text).
  • Measured relocation efficiency is about 93.5%. The simulation produced b = 159,461.05 USDT, x′ = 50.4741 WETH recovered flash capital, b′ = 157,262.60 USDT as the Stage 2 extraction input, y = 49.9515 WETH extraction repayment, and a′ = 9.3541 WETH delivered to EOA B. With η = a′/a = 9.3541/10.0000 ≈ 93.5%, the total loss is 6.5% (0.6459 WETH): roughly 1.2% in nominal protocol fees across the four 0.3% swaps, plus slippage and imbalance between the Stage 1 dislocation depth and the Stage 2 extraction volume.
  • Role separation is sufficient but not necessary. Even if the Principal is also the Operator (O = P), causal non-closure persists, so the failure mode is internal to composable execution rather than to intent delegation.
  • The failure modes are structurally generative, creating a detection asymmetry. The authors compare this to polymorphic or metamorphic malware: adversaries can mutate the observable transfer-layer structure while preserving identical net economic migration, and the space of possible execution graphs expands combinatorially as routing layers and the MEV supply chain evolve.
  • The boundary conditions are explicit. Transfer-layer monitoring is judged structurally insufficient where three conditions coincide: modular composition, open participation under market incentives, and invariant-driven state dynamics.
  • Not reported: the paper presents no dataset sizes, no trained models or benchmark comparisons against commercial analytics platforms, and no false-positive or detection-recall figures for existing tools. Its quantitative evidence is the single case-study transaction and the single simulation run.

Methodology in Plain English

The authors build their argument in four layers. First, they dissect one real Mainnet transaction and hand-label the three roles: who signed the order, who submitted the transaction, and who received the output. Second, they write down a deliberately generous model of what a transfer-layer observer can see — a graph of addresses and token-transfer edges, plus transaction metadata — and then define precisely what it would mean for a value migration to be "transfer-recoverable": there must be a directed path that uniquely and deterministically represents the causal link between the sender's loss and the receiver's gain. Uniqueness is the key word, because many benign routes through shared contracts could produce the same edges. Third, they derive two failure propositions from that model and combine them into an informal incompleteness theorem. Fourth, rather than stopping at the real case, they construct a general attack pattern using two constant-product AMM pools and work out the algebra of the four swaps in a zero-fee setting, then test it empirically by replaying it against a forked copy of Ethereum state with real pool reserves and real 0.3% fees, measuring how much of the principal survives the round trip. No machine learning, labelled datasets or classifier benchmarks are used; the evidence is analytical plus a single executable simulation.

Why This Matters

Impact on research. The paper reframes a practical complaint ("graph tracing misses things") as a formal observational limit, and it does so without assuming protocol bugs or exotic adversaries. That changes the burden of proof for transfer-graph learning: improvements in motif classification, taint propagation or GCN-style node classification cannot recover information that the abstraction discards. It also connects AML research to execution-trace work such as TxSpector and MFTracer and to MEV research on ordering-dependent value channels, suggesting where a richer observational basis already exists. The finding that the same transfer-layer motifs can arise from benign arbitrage and from deliberate relocation also explains the persistent distribution shift that graph-based detectors face.

Real-world applications:

  • Blockchain analytics vendors (the paper names Elliptic, Chainalysis and TRM Labs) could use the two failure types as test cases for coverage gaps, and as motivation for execution-semantic features rather than transfer-only heuristics and entity labels.
  • VASPs and compliance teams subject to FATF Recommendation 15, the Travel Rule and EU Regulation 2023/1113 could reassess what an exposure score actually certifies when the underlying transaction was filled by a third-party executor.
  • Law enforcement and incident responders tracing stolen or phished funds — the case study's principal is phishing-labelled — get a concrete explanation of why funds can appear to "leap" from a victim address to a fresh EOA with no intervening transfer edge.
  • DeFi protocol and audit teams (the paper cites audit findings that MEV-protected designs can introduce new opacity, for example internal PMM-to-user settlements that bypass indexed swap events) get a checklist of architectural features that defeat downstream forensics.

Industry relevance. The authors describe transfer-graph AML as the bedrock of current industrial forensics and argue it has reached an observational boundary under modular composition and open participation markets. The practical implication is that compliance tooling may need to ingest call graphs and internal state transitions, not just token transfers — a substantial change to data pipelines, evidence standards and how suspicious-activity narratives are written.

Future Directions

  1. Build execution-semantic monitoring. The paper's headline recommendation is to reconstruct economic causality from atomic execution logic, call graphs and state invariants rather than graph connectivity, mirroring the shift in security from signature-based to behavioural detection. How to do this at chain scale is left open.
  2. Resolve attribution ambiguity at the execution layer. Once execution semantics are visible, intents must still be disentangled from mechanical execution — separating the party who wanted the trade from the searcher who happened to submit it — which the authors flag as an unsolved problem introduced by their own proposal.
  3. Model economic intent inference. Identifying off-chain coordination and signalling (order signing, solver behaviour, pre-arranged settlement) is identified as a required modelling capability, and the paper notes that key identity and coordination signals often reside off-chain or in intent-matching layers.
  4. Formalise the integration of execution semantics into AML attribution. The conclusion calls for formal work linking the two, and the gap between the informal Theorem 4.1 and a fully general result — including whether the reconstruction can be made efficient, not merely possible — remains open. A larger-scale empirical study than the single case study and single 10 WETH simulation would also be needed to gauge how prevalent these patterns are in practice.

Target Audience

Compliance and investigations staff at VASPs and analytics vendors who need to understand where exposure scoring breaks down; blockchain-forensics researchers and PhD students working on transaction-graph learning or taint analysis; DeFi protocol designers and auditors reasoning about how intent-centric and MEV-aware architectures affect post-hoc attribution; and policy or supervisory technologists assessing whether the Travel Rule and VASP-centric frameworks still have identifiable intermediaries to attach to in permissionless settings. Readers wanting empirical detection benchmarks rather than a structural argument should look elsewhere — this paper offers a case study, a construction and a formal boundary, not a comparative evaluation of tools.

Authors’ abstract

Transfer-based anti-money laundering (AML) systems monitor token flows through transaction-graph abstractions, implicitly assuming that economically meaningful value migration is sufficiently encoded in transfer-layer connectivity. In this paper, we demonstrate that this assumption, the bedrock of current industrial forensics, fundamentally collapses in composable smart-contract ecosystems. We formalize two structural mechanisms that undermine the completeness of…

The full abstract on arXiv
Read the original paper