Skip to content
Crypto.world

Research

ConneX: Automatically Resolving Transaction Opacity of Cross-Chain Bridges for Security Analysis

Cross-chain bridges move assets between blockchains, but they deliberately break the visible link between the transaction on the source chain and the matching transaction on the destination chain — a…

ConneX: Automatically Resolving Transaction Opacity of Cross-Chain Bridges for Security Analysis
arXiv
2511.01393
Published
2025-11-03
Authors
Hanzhong Liang, Yue Duan, Xing Su, Xiao Li, Yating Liu, Yulong Tian, Fengyuan Xu, Sheng Zhong

AI summary

Overview

Research area: Blockchain security and cross-chain interoperability — specifically cross-chain transaction pairing and fund-flow traceability.

Technical level: Advanced. The paper combines formal problem definitions (fields, transaction instances, quintuples, pruning formulas) with an LLM-driven semantic inference pipeline and a validation algorithm, so comfort with smart contracts, EVM event logs, and ABI decoding helps.

Scope (one sentence): The paper presents ConneX, a system that automatically reconstructs the hidden pairing between source-chain and destination-chain transactions for cross-chain bridges, without cooperation from the bridge operators, and demonstrates its use for security analysis including illicit fund tracing.

What This Paper Is About

Cross-chain bridges move assets between blockchains, but they deliberately break the visible link between the transaction on the source chain and the matching transaction on the destination chain — a problem the authors call transaction opacity. This makes it hard to trace stolen funds, detect vulnerabilities, or build transaction graphs across chains. The paper's goal is to automatically re-establish those source–destination transaction pairs using only public on-chain data, by identifying the same semantic information (recipient, chain, amount, asset, timestamp) on both sides of the bridge.

Key Contributions

  1. First automated, generalized cross-chain pairing mechanism. The authors present ConneX, which they describe as the first automated and generalized discovery mechanism for hidden cross-chain transaction pairs, operating without any collaboration from black-boxed bridge platforms. A prototype is referenced at an anonymous repository link.
  2. A "semantic quintuple" identifier. They define a general five-element identifier — destination address, counterpart chain, asset type, amount, and timestamp — as a robust way to characterize a cross-chain transaction and match it across chains.
  3. Semantic-aware key-value dual pruning. They propose a two-stage pruning method: an LLM filters a huge space of candidate data fields down to plausible ones, then an examiner module validates those candidates against actual transaction values.
  4. Extensive real-world evaluation and security application. They evaluate on roughly 500,000 transactions from five bridges and apply ConneX to two real-world hacking incidents, including tracing funds from the Bybit Hack.

Main Findings

  • High pairing accuracy: On a dataset of approximately 500,000 transactions from five major bridge platforms (Stargate, DLN, Multi, Celer, Poly) collected between February 2021 and March 2024, ConneX achieves an average F1 score of 0.9746.
  • Clear margin over baselines: This outperforms the baselines by at least 20.05%. The baselines compared are heuristic (B_rule), similarity-based (B_sim), and hybrid (B_hybrid) approaches.
  • Massive search-space reduction: The core pruning method cuts the semantic search space from over 10^10 candidates to fewer than 100.
  • Practical efficiency: Average processing time is 0.4 seconds per transaction, a speedup of up to 9x compared to the baseline.
  • Breadth of applicability: Among the top 10 cross-chain bridges ranked by Total Value Locked (as retrieved from Chainspot in May 2024), 7 are compatible with the paper's generalized cross-chain model.
  • Concrete illustration of the search problem: A single motivating transaction with 9 event logs yields 144 distinct fields, giving C(144,5) = 481,008,528 possible five-field combinations for just that one transaction.
  • Real-world tracing success: Combined with existing fund-tracing techniques, ConneX identified hacker-related cross-chain transfers and flagged intermediary addresses on destination chains; in the Bybit Hack it identified a transit address receiving $1 million USDCs on Solana.

Methodology in Plain English

The system works in five steps:

  1. Decode. Raw transaction data and event logs are decoded using the smart contract's ABI, turning them into nested key-value structures. Each data point gets a unique "field" name built from the path of keys leading to it, which avoids name collisions across nested logs.
  2. Group. Transactions are bucketed into categories that share exactly the same set of fields. This collapses many individual transactions into a small number of structural templates, so the expensive reasoning only has to be done once per template.
  3. Ask an LLM for candidates. For each category, an LLM is shown a few sample transactions and asked, with a worked example and chain-of-thought guidance, to propose which fields might correspond to each of the five quintuple elements, returning a JSON list with confidence scores. This is where the enormous combinatorial space is cut down quickly.
  4. Verify with real values. An "examiner" module then checks the LLM's guesses against what actually happened on-chain. It analyzes the asset flow in the source transaction to confirm which candidate amount and token fields are real; it looks for matching destination transactions within a time window to confirm the destination, chain, and timestamp fields; and it removes ambiguous candidates using consistency (do candidates always extract the same value?) and uniqueness (is the value variable enough to distinguish transactions?).
  5. Pair. With a validated quintuple per category, the system matches transactions across the two chains and outputs the pairing relationships.

The reversal between source and destination chains is handled by simply swapping the roles of the two transaction sets.

Why This Matters

  • For research: Cross-chain attack detection, cross-chain monitoring, and contract vulnerability analysis all assume the pairing relationships that ConneX produces. Prior tools were rule-based and covered only a handful of bridges; this work claims the first generalized, automated approach and supplies the missing data layer for downstream security research.
  • Real-world applications:
    • Illicit fund tracing: following stolen assets through bridges instead of stopping at the bridge boundary, as demonstrated with the Bybit Hack.
    • Anti-money-laundering analysis: building cross-chain fund-flow graphs that money-laundering studies currently cannot complete.
    • Cross-chain MEV and sandwich attack analysis: studying value extraction that spans multiple chains.
    • Cross-chain vulnerability detection and behavioral modeling: enabling graph-based analyses and deeper contract-semantics understanding that require linked cross-chain records.
  • Industry relevance: The system needs no cooperation from bridge operators and no knowledge of the off-chain relayer logic, so exchanges, analytics firms, and incident responders can apply it to real bridges using only public data.

Future Directions

  • Extending beyond EVM-compatible public chains. The paper explicitly limits scope to public, EVM-compatible blockchains and excludes non-EVM chains such as Bitcoin and private networks; generalizing the decoding and matching logic to those environments remains open.
  • Covering excluded bridge types. Bridges that do not operate via smart contracts (for example, the Avalanche Bridge) and canonical Layer 2 native bridges with their own security and verification models are out of scope and would need separate treatment.
  • Handling bridges with non-public or opaque records. The paper notes that some bridges prioritize privacy or security and produce inaccessible transaction records, which the current approach cannot address.
  • Scaling and hardening the LLM component. The approach depends on an LLM for candidate generation and reports comparisons across different LLM backends; robustness of prompts, sensitivity to the sampled transactions per category, and behavior as new bridge naming conventions appear are natural next questions.

Target Audience

Blockchain security researchers and graduate students working on cross-chain systems; security engineers and incident responders at exchanges, analytics companies, and audit firms who need to trace funds across bridges; and bridge developers or auditors interested in how their transaction semantics appear to external analysis. Readers need some familiarity with EVM transactions, event logs, and ABIs to follow the design details.

Authors’ abstract

As the Web3 ecosystem evolves toward a multi-chain architecture, cross-chain bridges have become critical infrastructure for enabling interoperability between diverse blockchain networks. However, while connecting isolated blockchains, the lack of cross-chain transaction pairing records introduces significant challenges for security analysis like cross-chain fund tracing, advanced vulnerability detection, and transaction graph-based analysis. To address this gap, we…

The full abstract on arXiv
Read the original paper