Skip to content
Crypto.world

Research

FlowShield: cryptocurrency anti-money laundering with transaction semantics parsing and fund flow tracking

Modern launderers break stolen assets into many fragments within minutes, chaining direct transfers, token swaps, indirect transfers, and cross-chain bridge deposits across multiple blockchains.

FlowShield: cryptocurrency anti-money laundering with transaction semantics parsing and fund flow tracking
arXiv
2608.17355
Published
2026-08-18
Authors
Qishuang Fu, Andreas Deppeler, Joseph K. Liu, Yixin Liu, Shirui Pan, Qin Wang, Weiqing Wang, Tsz Hon Yuen

AI summary

Overview

Research area: Cryptocurrency anti-money laundering (Crypto AML) — specifically transaction-level detection of laundering behavior on account-based and UTXO-based blockchains, combined with automated investigator-facing reporting.

Technical level: Advanced. The paper assumes familiarity with graph neural networks, large language model fine-tuning (LoRA), and blockchain transaction anatomy (external/internal/token transfers, bridges, swaps).

Scope: FlowShield is a four-module framework that parses transaction semantics, reconstructs localized fund-flow subgraphs, fuses LLM-encoded text with GCN-encoded structure, and outputs both laundering labels and suspicious activity reports (SARs), evaluated on four real-world laundering datasets.

What This Paper Is About

Modern launderers break stolen assets into many fragments within minutes, chaining direct transfers, token swaps, indirect transfers, and cross-chain bridge deposits across multiple blockchains. Existing detectors treat transactions as isolated samples, rely on graph topology alone, or abstract away what a transaction actually does — and they output a bare label rather than something an investigator can act on. FlowShield aims to recover behavior-level transaction semantics, explicitly trace how value propagates from one transaction to the next, and turn flagged multi-hop flows into readable reports.

Key Contributions

  1. A new modeling perspective. The authors state this is the first work to jointly model transaction semantics and fund-flow propagation for Crypto AML, moving beyond address- or transaction-level classification to modeling how funds propagate through diverse transaction behaviors.

  2. A unified detection and reporting framework. FlowShield comprises four modules — semantics parsing, fund-flow construction, text–structure fusion, and detection plus report generation — that combine textual behavioral descriptions with graph-structural context and organize detected multi-hop flows into SARs.

  3. The first public multi-chain AML dataset. BybitML is constructed from the 2025 Bybit Hack, covering two blockchains (Ethereum and Bitcoin) with transaction-level semantic annotations and laundering labels. It is released at an OSF link given in the paper; the FlowShield implementation is also released.

  4. Extensive experimental validation. Evaluation on BybitML plus two public laundering datasets (Upbit, AscendEX) against thirteen representative baselines, with the claim of consistently best detection performance and robustness across heterogeneous laundering scenarios.

Main Findings

  • Best F1 on all four datasets. FlowShield scores 97.2% (Bybit-EC), 97.4% (Bybit-BC), 98.1% (Upbit), and 99.2% (AscendEX), giving an average rank of 1.0 across the four datasets and an average F1 of 98.0% as stated in the abstract.

  • Multimodal graph baselines are the closest competitors. GraphGPS reaches 96.0 / 94.7 / 96.5 / 97.6 and GraphFormers 95.7 / 94.4 / 96.4 / 97.5 across Bybit-EC / Bybit-BC / Upbit / AscendEX, ranking second (2.3) and third (3.3) respectively.

  • Heuristics and shallow ML are unstable. Cyclic scores 6.9 on Bybit-EC and 0.7 on Bybit-BC; MLP scores 9.2 on Bybit-EC; XBlock reaches 16.7 on Bybit-EC. XBlock and Tracer cannot run on Bitcoin, so they cannot detect Bybit-BC, and XBlock is excluded from Upbit and AscendEX because those labels were collected using XBlock.

  • Day-by-day stability is the strongest claim. On the 14 daily slices of Bybit-EC, FlowShield achieves an average rank of 1.1 and ranks first on 12 of 14 days. Many baselines collapse late in the window: LR, SVM drop to 0.0 on Days 12–14, while FlowShield records 93.1, 90.2, and 93.1 on those days.

  • Label reliability was manually checked. 800 randomly sampled transactions were audited — 200 laundering and 200 normal from each chain — and all audited labels matched their assigned labels.

  • Dataset scale differs sharply across chains and incidents. Bybit-EC has 58,790 addresses and 2,616,736 transactions at a 7.6% laundering ratio; Bybit-BC has 62,409 addresses and 134,280 transactions at 23.2%; Upbit has 16,397 addresses, 40,287 transactions, 43.4% laundering; AscendEX has 6,713 addresses, 29,074 transactions, 6.7% laundering.

  • SAR and efficiency results are referenced but not reported in the supplied text. The paper poses RQ2 (ablation), RQ3 (hyperparameter and semantic-parsing drift robustness), and RQ4 (scalability), and includes Figures 8 (parameter sensitivity) and 9 (SAR quality, detection time), but the numerical outcomes of those analyses are cut off in the available content.

Methodology in Plain English

Step 1 — Work out what each transaction actually did. Raw on-chain fields only show sender, receiver, and amount, so two transactions can look identical structurally while doing completely different things. FlowShield groups all transfer records sharing one transaction hash (external, internal, and token transfers) and inspects the relations inside that group. If the sender sends one token and receives a different one back within the group, it is a token swap. If the receiver forwards the same token in the same amount, it is an indirect transfer. If the receiver is a known bridge contract and the event log carries parseable destination-chain and recipient data, it is a cross-chain transfer. Anything else stays a direct transfer. Each type is then written out using a type-specific text template.

Step 2 — Reconstruct how the money moved. Using the whole neighborhood would drown the signal in noise, so FlowShield builds three small subgraphs around each target transaction: an upstream subgraph of transactions that funded the sender within a time window, a downstream subgraph of the earliest outflows from the receiver, and a parallel subgraph of similar-sized concurrent inflows. Upstream and downstream collection stops once cumulative value reaches a set fraction of the target amount. A rule set then classifies the movement into one of five patterns — single outflow, entire disperse, partial disperse, mix, or no forward — and verbalizes it into a "flow text."

Step 3 — Fuse two views. The semantic text and flow text are concatenated and encoded by a Llama3-8B model fine-tuned with LoRA on supervised labels (labels are targets only, never placed in the text); token embeddings are averaged with attention-mask mean pooling. Separately, a two-layer GCN runs over the transaction graph, and a transaction's structural vector is the concatenation of its sender's and receiver's node embeddings. A bidirectional attention layer lets each representation look at the other, and the two directional outputs are concatenated and normalized.

Step 4 — Detect and report. A two-layer MLP makes the laundering/normal decision. Flagged transactions are merged into connected multi-hop flows, drawn as a Sankey-style FlowDiagram, and GPT-4o is prompted with the aggregated transaction descriptions to produce a summary and a set of red flags.

Experimental protocol. 6:2:2 train/validation/test split, five repetitions with mean and standard deviation reported, F1 as the primary metric given class imbalance. Thresholds are τ_t = 24 (time window), τ_f = 0.9 (flow ratio), τ_a = 0.05 (amount similarity). The GCN has hidden dimension 16 producing 32-dimensional structural embeddings; the interaction layer uses 4 attention heads and dropout 0.1; the classifier has hidden dimension 64 and is trained with Adam at learning rate 1e-3, batch size 64, for 100 epochs. Text encoders are fine-tuned on 5,000 transactions sampled from the training split of each run. All experiments ran on Python 3.13.12 with one NVIDIA RTX 4090 and 46GB RAM.

Why This Matters

Research impact. The paper reframes Crypto AML from address classification into a problem of behavioral semantics plus value provenance, and it supplies a multi-chain benchmark where previously the field relied on single-chain, dated datasets such as Elliptic (2019) and EthereumHeist (2022). The comparison table makes the gap concrete: no prior method in the list covers all four transaction semantics together with laundering-flow reconstruction and investigator reporting.

Real-world applications:

  • Exchange and custodian compliance teams screening incoming deposits against detected laundering flows before funds are cashed out.
  • Incident response after a hack. BybitML is built from the 2025 Bybit Hack, so the pipeline is directly applicable to post-breach fund tracing during the critical first hours.
  • Cross-chain investigation. Bridge-log parsing and cross-chain transfer semantics address the specific case where funds leave the current chain and single-chain graphs go blind.
  • Regulatory and law-enforcement reporting. Auto-generated SARs with flow diagrams and red flags reduce the manual work of turning detections into filings.

Industry relevance. The combination of an open dataset, open implementation, and an LLM-based reporting layer targets a real operational bottleneck — analysts currently must hand-assemble the narrative around an alert. A 98.0% average F1 with stable day-by-day performance under sparse labels and class imbalance is the kind of result that matters for production screening, where late-window degradation (as seen in LR and SVM dropping to 0.0) is a practical failure mode.

Future Directions

  • Robustness to semantic-parsing drift. RQ3 explicitly asks whether FlowShield survives drift in the parsing rules, but the numerical answer is not visible in the supplied text. How the four hand-specified semantics hold up as new DeFi primitives appear is an open question.
  • Generalization beyond the studied incidents. All four datasets stem from three incidents (Bybit, Upbit, AscendEX), and Upbit/AscendEX labels were derived by XBlock's heuristic search. Testing on laundering typologies not represented in these incidents would clarify whether the semantic templates transfer.
  • Scaling to full-chain volumes. RQ4 addresses scalability and Figure 9c reports detection time, but the figures are not given here. Building upstream, downstream, and parallel subgraphs per transaction is inherently neighborhood expansion, so cost at production scale remains to be characterized.
  • Evaluating the reports themselves. Figure 9b reports SAR quality, but the evaluation protocol for generated summaries and red flags — and whether investigators trust and act on them — is not described in the available content.

Target Audience

Researchers and graduate students working on graph learning applied to blockchain forensics; AML and compliance engineers at exchanges, custodians, and analytics vendors; and incident responders who need to trace stolen funds across chains and produce defensible written reports. Readers should be comfortable with GNN architectures and LLM fine-tuning, since the paper does not introduce those building blocks.

Authors’ abstract

Cryptocurrency anti-money laundering (Crypto AML) is increasingly challenged by sophisticated laundering behaviors that rapidly fragment stolen assets through diverse semantics and across multiple blockchains. Existing Crypto AML methods often simplify transaction semantics, rely on topology-centric signals, or output isolated detection labels. In this paper, we present \textsc{FlowShield}, a Crypto AML framework for transaction-level laundering detection and…

The full abstract on arXiv
Read the original paper