Skip to content
Crypto.world

Research

Heterogeneous Cross-Chain Transaction Tracing for Solana Bridges via Candidate-Set Selective Decision

Tracing stolen or laundered crypto across chains is hard when the source chain and target chain record transactions in fundamentally different ways. Existing tracing tools assume EVM-style structured…

Heterogeneous Cross-Chain Transaction Tracing for Solana Bridges via Candidate-Set Selective Decision
arXiv
2609.11413
Published
2026-09-10
Authors
Wenjie Dou, Zheng Che, Meng Shen, Hanbiao Du, Qing Li, Yan Qiang

AI summary

Overview

Research area: Blockchain security and cross-chain forensics — specifically, tracing transactions that move funds from EVM-compatible chains onto Solana through cross-chain bridges.

Technical level: Advanced. The paper assumes familiarity with cross-chain bridge architectures, Solana's account model, cross-program invocations (CPIs), associated token accounts (ATAs), and machine-learning classification with a reject option.

Scope (one sentence): The paper formalizes four Solana-bound cross-chain transaction modes and introduces SolTracer, a system that maps heterogeneous EVM and Solana execution records into a shared semantic event space and then decides whether to accept, or explicitly abstain from, an association for each source transaction.

What This Paper Is About

Tracing stolen or laundered crypto across chains is hard when the source chain and target chain record transactions in fundamentally different ways. Existing tracing tools assume EVM-style structured event logs and assume that every source-chain transfer has exactly one identifiable counterpart on the destination chain — neither assumption holds for Solana, which has no standard event logs and where bridging can be delayed, batched, or fulfilled by third-party solvers. The paper's goal is to correlate EVM-to-Solana bridge transfers reliably while refusing to guess when no valid Solana target actually exists, and then to use those correlations to describe the real-world Solana bridge ecosystem.

Key Contributions

  1. A four-mode taxonomy of Solana-bound cross-chain transfers. The authors identify and formalize Message Redemption (MRT), Pool Settlement (PST), Solver Fulfillment (SFT), and Burn–Mint (BMT) transactions, characterizing each workflow and its specific observability gaps across four dimensions: recipient account representation, amount dynamics, asset identity, and authorization provenance.

  2. SolTracer, a candidate-set selective decision tracing method. The system has three modules — bridge protocol identification, semantic event reconstruction, and candidate selective association — the last of which adds a candidate-set-level reject option so the tracer can output an abstention symbol instead of forcing a top-ranked match.

  3. A benchmark against six state-of-the-art methods. AttrMatch, EntityMatch, MultiMatch, CONNECTOR, ABCTRACER, and ConneX are re-adapted to the heterogeneous EVM-to-Solana setting using a unified semantic-event representation, and SolTracer is reported to achieve the highest F1 scores in closed-world and open-world settings.

  4. An empirical study of the Solana bridge ecosystem. Using the recovered associations, the authors examine count–value divergence across bridge mechanisms, cross-asset shifts, target observability gaps, and mechanism-dependent association reliability.

Main Findings

  • Open-world performance gain: Under the open-world setting with a 50% TA (target-absent) ratio, SolTracer improves F1 by 20.16% over the strongest state-of-the-art baseline. Absolute precision, recall, and F1 values for this and the other scenarios are not included in the truncated content provided.

  • Best-in-class across three scenarios: The paper reports that SolTracer achieves the highest F1 scores in closed-world association, open-world association, and cross-source-chain generalization, across all six compared baselines.

  • Dataset scale: Acquisition pipelines retrieved 38,317 bridge records across the main and generalization collection rounds and retained 7,082 verified source–target pairs after deduplication, dual-chain verification, and bridge-specific consistency checks. Augmenting each correlated pair with 30 protocol-compatible non-correlated pairs yields 212,460 NCPs and 219,542 candidate pairs in total.

  • Per-bridge verified pairs: Wormhole (MRT) 1,625 pairs from 5,977 raw records; Allbridge (PST) 1,341 pairs from 1,480 raw records; deBridge (SFT) 1,597 pairs from 1,600 raw records; Circle CCTP (BMT) 2,519 pairs from 29,260 raw records.

  • Threshold sensitivity: A grid sweep of the candidate-pair threshold τ_p from 0 to 1 in steps of 0.01 over 12,400 validation pairs (371 positive, 12,029 negative) found an optimal τ_p = 0.94; recall stays above 0.92 until τ_p = 0.98, while precision rises monotonically with the threshold.

  • Heterogeneity is the core obstacle: Because Solana lacks standard event logs and fragments bridge semantics across invoked programs, account roles, inner instructions, and balance deltas, uniform EVM-tailored parsers cannot align source and target fields directly.

  • Abstention is necessary, not optional: Asynchronous processing, cross-asset solver settlement, and unexecuted transfers break the one-to-one source–target assumption, so ranking without a reject option systematically selects the top-scoring distractor.

  • Empirical ecosystem findings: The analysis reports a stark count–value divergence across bridge mechanisms, a prevalence of cross-asset shifts, and a decoupling between on-chain settlement and explorer visibility.

  • Concrete illicit-use motivation: The paper cites Solana drainer campaigns that bridged over one million USD of stolen assets from Solana to Ethereum via AllBridge, and the BingX exploit, where attackers routed funds onto Solana via deBridge and performed token swaps to disrupt asset lineage.

Methodology in Plain English

The researchers start by treating the task as a matching problem with an escape hatch: given a source-chain transaction, collect a limited set of plausible Solana candidates, rank them, and then decide as a set whether any of them is trustworthy enough to accept.

Step 1 — Identify the bridge protocol. For each source transaction, the system scores how well the transaction's routing cues match known bridge specifications (router and program identifiers, event topics, route directions, asset mappings, expected latency windows, participant roles). If the best match falls below a threshold τ_proto, the system abstains immediately rather than emitting noise. This is an admission-control step that keeps unsupported transfers out of candidate construction.

Step 2 — Rebuild the transaction semantics. Each protocol is tied to one of the four transaction modes, and each mode has a dedicated adapter that converts raw execution data — EVM receipts and logs on one side, Solana inner instruction trees, account roles, and balance deltas on the other — into one common record format holding the route domain, participant roles, a normalized asset identifier, a decimal-adjusted amount, and provenance metadata. For example, the Message Redemption adapter unrolls inner CPIs to the SPL Token program and resolves the associated token account back to the recipient, while the Pool Settlement adapter uses net balance deltas so pool fees and slippage do not break the amount comparison.

Step 3 — Build and filter the candidate set. Candidates are retained only if route, time window, expected or mapped asset, mode-specific amount relation, and recovered participant roles are compatible — allowing fee-adjusted pool payouts and cross-asset solver fulfillment rather than demanding exact equality. An empty set counts as an abstention, not a ranking failure.

Step 4 — Rank, then decide whether to trust the ranking. A random forest over 14 semantic features (300 trees, balanced class weights, minimum leaf size two) ranks the candidates. Separately, a logistic model looks at set-level evidence — top score, score separation, time and amount consistency, asset/recipient/role consistency — and estimates whether the candidate set supports a valid association. If that probability clears the operating threshold τ, the top-ranked candidate is returned; otherwise the system abstains, and it records whether the abstention came from an empty candidate space or from a candidate set that failed the reliability test. Candidate count and minimum time distance are deliberately excluded from the rejector to avoid structural leakage.

Evaluation design: Ground-truth pairs are labelled Target-Present (TP), and Target-Absent (TA) groups are created by masking out the verified target. Data is split at the candidate-group level into training, validation, and test sets. Closed-world and open-world experiments use Ethereum-to-Solana data; generalization trains and calibrates on Ethereum and evaluates on Arbitrum, Base, and Polygon PoS. Main experiments are repeated with five fixed seeds (7, 11, 19, 23, 42) and means are reported. Beyond standard precision, recall, and F1, the paper defines group-level metrics P_g, R_g, and F1_g to measure the ability to abstain correctly when no target exists. Experiments ran on a Windows 11 64-bit laptop with an AMD Ryzen 9 5900HX 8-core CPU, an NVIDIA GeForce RTX 3060 Laptop GPU with 6 GB memory, and 16 GB RAM, with programs in Python 3.13.10.

Why This Matters

Impact on research: The paper breaks the field's implicit assumption that cross-chain tracing means EVM-to-EVM tracing with ABI-decoded logs on both sides. By formalizing Solana-bound modes and publishing a labelled benchmark spanning four source chains and four bridges, it gives subsequent work a shared problem statement, a distractor-generation protocol, and a reject-aware evaluation metric that penalizes confident wrong answers rather than rewarding coverage alone.

Real-world applications:

  • Exchange and custodian compliance: Screening inbound Solana deposits to check whether they trace back to flagged source-chain addresses.
  • Law-enforcement and incident response: Following funds after drainers and bridge exploits, where attackers deliberately hop chains and swap assets to break lineage.
  • Bridge risk monitoring: Quantifying which bridge mechanisms produce the most value per transaction and how often delivered assets differ from deposited ones.
  • Block explorer and analytics products: Reconstructing the recipient and amount that top-level Solana instruction parsing hides inside inner CPIs.

Industry relevance: The finding that on-chain settlement and explorer visibility are decoupled is directly operationally significant — any compliance pipeline built on explorer-displayed data alone inherits that blind spot. The paper also shows that a large share of real transfers follow cross-asset routes, which invalidates the simple asset-equality heuristics that commodity tracing tools typically apply.

Future Directions

  • Extending beyond Solana-bound flows: The taxonomy and adapters are built around EVM-to-Solana directionality; whether the same semantic-event approach transfers to other non-EVM targets with different account and execution models is open.
  • Protocol coverage and maintenance: Protocol specifications are described as extendable through expert guidance and semi-automated learning from documentation and verified transfers, but how quickly the registry degrades as bridges upgrade their programs is not evaluated in the presented content.
  • Filling out the reported results: The truncated content omits the absolute precision, recall, and F1 tables for closed-world, open-world, and generalization, as well as the ablation results and the learned rejector threshold τ — these are the natural targets for replication and critical comparison.
  • Adversarial robustness: Attackers who know tracing relies on intent logs, solver traces, and attestations could deliberately route through mechanisms that minimize those artifacts; testing whether the abstention behavior degrades safely under adversarial bridge selection is a logical next question.

Target Audience

Researchers and graduate students in blockchain security, financial forensics, and applied machine learning working on transaction-graph analysis or cross-chain data. Also relevant to blockchain-analytics engineers, compliance and investigations teams at exchanges and custodians, and bridge operators who need to understand how their settlement design affects downstream auditability. Readers without background in Solana's account model and CPI mechanics will find the methodology section demanding, though the four-mode taxonomy is accessible on its own terms.

Authors’ abstract

Solana is a rapidly growing high-throughput blockchain platform that has attracted substantial liquidity and user activity. However, this expansion has also drawn the attention of illicit actors, who frequently leverage cross-chain bridges to route illicit funds onto Solana to obfuscate transaction lineage. Unlike EVM-compatible platforms, Solana features distinct execution dynamics and lacks standard event logs, creating severe semantic gaps that prevent existing tracing…

The full abstract on arXiv
Read the original paper