# UniDetect: LLM-Driven Universal Fraud Detection across Heterogeneous Blockchains

Crypto.world

> Existing cryptocurrency fraud detectors are built for one blockchain at a time and depend on hand-crafted numerical features tailored to that chain, so they do not transfer when illicit funds move…

**arXiv**: 2604.12329

**Published**: 2026-04-14

**Authors**: Shuyi Miao, Wangjie Qiu, Shengda Zhuo, Fei Shen, Dan Lin, Xingtong Yu, Chua Tat-Seng, Zhiming Zheng

## AI summary

## Overview

- **Research area:** Blockchain security and financial fraud detection, specifically cryptocurrency account fraud detection across heterogeneous blockchains using large language models (LLMs) combined with graph neural networks.
- **Technical level:** Advanced. The paper assumes familiarity with graph representation learning (GCN, GAT, GraphSAGE), graph anomaly detection, LLM fine-tuning via LoRA, and reinforcement learning with REINFORCE.
- **Scope:** The paper proposes UniDetect, a multi-chain fraud account detection method that uses LLM-generated transaction summaries as node features for a dual-path graph neural network trained through a two-stage alternating strategy, evaluated on Ethereum, two Bitcoin datasets, and the Instagram social network.

## What This Paper Is About

Existing cryptocurrency fraud detectors are built for one blockchain at a time and depend on hand-crafted numerical features tailored to that chain, so they do not transfer when illicit funds move across chains. The paper's goal is a single detection framework that works on blockchains with fundamentally different data structures by having an LLM read raw transaction histories and write plain-language account summaries, which then serve as evidence for a graph model to classify accounts as fraudulent or benign. The authors motivate this with industry figures cited in the paper: fraud losses above USD 2.47 billion in the first half of 2025 (CertiK), and cross-chain criminal activity exceeding USD 21.8 billion in 2025, with nearly 20% of complex illicit cases spanning more than ten blockchains (Elliptic).

## Key Contributions

1. **First LLM-based universal detector across heterogeneous chains.** The authors present UniDetect as the first attempt to combine LLMs with customized fine-tuning to identify fraudulent accounts across blockchains with different data structures, rather than designing per-chain features.
2. **Transaction summaries as graph node features.** A forensic analysis agent uses domain knowledge and chain-of-thought prompting to convert each account's transactions into a natural-language summary; these summaries become the semantic features that the transaction-graph model learns from.
3. **Two-stage alternating training with RL.** Discriminative and residual summary analyst agents share an LLM backbone and are fine-tuned (via LoRA and sequence-level REINFORCE) in an outer loop, while the Dual-PGNN graph encoder is trained in an inner loop, jointly improving summary quality and downstream graph reasoning.
4. **Efficient graph construction.** Semantics-aware top-K neighbour sampling plus a structural-importance graph compression (SIGC) algorithm keeps subgraph size bounded so LLM and graph processing remain tractable.

## Main Findings

- **Strongest results on Ethereum:** UniDetect reaches Precision 91.39, Recall 93.95, F1 92.65, AUC 94.82, and KS 68.94 (percent), corresponding to improvements of +8.71, +5.40, +8.56, +12.15, and +7.58 over the best competing baselines.
- **Bitcoin-M results:** Precision 90.61, Recall 95.47, F1 92.98, AUC 93.83, KS 70.45, i.e. improvements of +1.78, +7.52, +5.41, +6.47, and +5.57.
- **Bitcoin-L results:** Precision 90.61, Recall 96.43, F1 93.41, AUC 95.22, KS 72.25, i.e. improvements of +1.03, +6.32, +3.61, +5.17, and +6.49.
- **KS gains across chains:** The abstract reports gains of 5.57 to 7.58 percent in Kolmogorov–Smirnov over existing methods, and the contribution list reports F1 gains of up to 8.56 percent.
- **Zero-shot cross-chain detection:** UniDetect trained on Ethereum and applied directly to Bitcoin identifies over 94.58% of fraudulent accounts. The paper describes this as asymmetric: strong at catching fraud, but with insufficient precision on normal accounts even after varying the classification threshold.
- **Cross-domain transfer:** On the Instagram graph dataset, where the minority business-user class is 36.29%, UniDetect generalizes well, with the abstract reporting a 6.06% F1 improvement over existing methods and the contribution list reporting a 4.47% AUC improvement on cross-domain datasets.
- **Zero-shot LLM variant still competitive:** UniDetect† (no fine-tuning) already scores F1 85.38 and KS 63.32 on Ethereum and F1 89.74 / KS 67.41 on Bitcoin-M, beating several trained baselines but trailing the full model.
- **Ablation results on Ethereum (F1 / AUC):** removing two-stage training drops to 85.65 / 82.52; removing the residual agent drops to 78.37 / 79.81; removing RL drops to 87.50 / 86.63; removing Dual-PGNN drops to 88.50 / 89.63, versus 92.65 / 94.82 for the full model.
- **Compression trades accuracy for speed:** removing graph compression ("w/o GC") slightly raises F1 to 93.62 and AUC to 94.14 but costs 40.37 seconds per account versus 1.26 seconds per account for the full model.
- **Efficiency:** Inference latency is 1.47 ms/sample at 680 samples/s throughput; two-stage training takes roughly 13 hours; LoRA keeps trainable parameters at only 0.07% of the total (the sentence is cut off in the supplied text).
- **Hyperparameter stability:** Performance stays competitive across a broad range of settings for subgraph size N_c, the structural weight β, and the tri-view loss weights λ1 and λ2, first improving and then declining as λ1 and λ2 grow.
- **Baselines struggled:** general graph learning and graph anomaly detection methods are limited by severe label imbalance and thin transaction information, single-chain methods by manual feature engineering, and multi-chain methods by separating sequence encoding from graph reasoning.

## Methodology in Plain English

The pipeline has five steps.

**1. Shrink the graph around each account.** Processing an entire blockchain graph is too expensive, so for each labelled account the method samples its neighbours hop by hop (2 hops, up to K = 10 neighbours per hop), preferring neighbours with high average transaction value and breaking ties by total value. A second step, structural importance graph compression, scores each neighbour by a combination of its transaction amounts, its in/out degree, and its shortest-path distance from the central account, keeps the top N_c = 10, and re-adds all nodes on the shortest paths to those kept nodes so connectivity survives.

**2. Have an LLM write a summary of the account.** A forensic analysis agent, built on gemma-3-4b-it with chain-of-thought prompting, reads the account's transactions and produces a natural-language summary describing its behaviour. The authors argue this is what makes cross-chain operation possible: the LLM absorbs the differences in data format instead of engineers writing per-chain features.

**3. Split the summary into useful and useless parts.** Two analyst agents share the same LLM backbone. The discriminative agent keeps detection-relevant content organized around transaction patterns, fund flows, associated addresses, and temporal signs. The residual agent pulls out the rest: hallucinated or over-reasoned text and purely factual statements that carry no risk signal. A frozen Sentence-BERT model embeds both parts, and a third embedding is made from the original summary.

**4. Train in two alternating stages.** In the outer stage the LLM analysts are fine-tuned with LoRA while the graph encoder is frozen; the loss on the graph side is turned into a reward and used with REINFORCE plus a moving-average baseline to push the analysts toward better summaries. In the inner stage the LLM is frozen and the graph encoder trains using the improved summaries. The graph encoder, Dual-PGNN, has two branches that each combine a GNN over the subgraph with a linear projection that preserves the account's own summary, then fuses the branches with attention. The training objective has three terms: binary cross-entropy on the discriminative branch, a KL divergence pushing the residual branch toward a uniform prediction (so it carries no class signal), and an orthogonality penalty keeping the two representations from interfering.

**5. Predict.** At inference, the original summary and the discriminative summary are fed into the two Dual-PGNN branches, and a fully connected layer outputs the fraud probability.

To prevent the LLM from exploiting memorized labels, the authors state that account identifiers and label information are excluded from all inputs and prompts.

## Why This Matters

**Research impact.** The paper reframes blockchain fraud detection from chain-specific feature engineering to LLM-mediated semantic representation, and shows that multimodal fusion (text plus graph structure) can be trained jointly rather than in separate stages. The two-stage alternating scheme with RL feedback from a downstream graph task is a reusable template for coupling LLMs to structured-data models under compute constraints.

**Real-world applications:**

- Exchange and custodian compliance teams screening depositors across multiple chains with one model rather than one per asset.
- Cross-chain bridge and DeFi protocol monitoring, where funds are reorganized into uniform liquid assets to evade single-chain surveillance.
- Blockchain analytics vendors that need to onboard a new chain quickly without rebuilding feature pipelines.
- Law-enforcement and regulatory forensics, where natural-language account summaries double as human-readable evidence for investigators.

**Industry relevance.** The reported inference cost of 1.47 ms/sample at 680 samples/s and a one-time ~13 hour training run suggest deployment is feasible, and the 0.07% trainable-parameter figure indicates the LLM adaptation is cheap relative to full fine-tuning. The paper also speaks directly to a documented regulatory gap: monitoring designed around individual blockchains is exactly what cross-chain laundering exploits.

## Future Directions

- **Fixing the precision asymmetry in zero-shot transfer.** The paper reports that cross-chain zero-shot inference catches over 94.58% of fraudulent accounts but is imprecise on normal accounts; calibrating or adapting the decision threshold per target chain is an open problem.
- **Extending beyond the two chains tested.** Evaluation covers Ethereum and two Bitcoin datasets; whether the forensic agent handles chains with very different execution models (e.g., account-based versus UTXO versus smart-contract-centric ecosystems) is not established in the reported results.
- **Reducing the LLM cost.** Summary generation runs per account, training takes roughly 13 hours, and removing compression raised per-account processing to 40.37 seconds; caching, distillation into smaller models, or batch summarization would improve scalability to the largest graphs (Bitcoin-L has 20,085,231 nodes).
- **Making the summaries auditable.** The paper positions summaries as evidence for fraud decisions but does not report a human evaluation of summary faithfulness by investigators, which would be a natural next validation step for regulated deployment.

## Target Audience

Researchers and graduate students in graph machine learning, LLM applications, and security/privacy who are interested in multimodal fusion and LLM-to-GNN coupling. It is also aimed at practitioners in blockchain analytics, exchange compliance, and regtech who need cross-chain detection that does not require per-chain feature engineering. Readers without background in graph neural networks or reinforcement learning will find the method sections dense, since the paper assumes familiarity with CARE-GNN-style backbones, LoRA, and policy-gradient fine-tuning.

## Authors’ abstract

As cross-chain interoperability advances, decentralized finance (DeFi) protocols enable illicit funds to be reorganized into uniform liquid assets that flow throughout the cryptocurrency market. Such operations can bypass monitoring targeted at individual blockchains and thereby weaken current regulatory frameworks. Motivated by these, we introduce UniDetect, a multi-chain cryptocurrency fraud account detection method based on large language models (LLMs). Specifically, we…

[The full abstract on arXiv](https://arxiv.org/abs/2604.12329)

[Read the original paper](https://arxiv.org/abs/2604.12329)

---

Canonical: https://crypto.world/paper/unidetect-llm-driven-universal-fraud-detection-across-heterogeneous-2604.12329
