DeFiFusion: Combining Transaction Events with Smart Contracts to Detect Price Manipulation Attacks
Rui Cao, Shaojing Fan, Liming Fang, Yuchan Liu, Yingying Jiao, Zhenguang Liu
cs.CR · cs.AI · cs.SE
Quick verdict
Claims dual-modal event+contract fusion: 98.67% recall, 96.10% precision on 225 DeFi attacks; strong, but curated data and reworked baselines warrant caution.
Summary
DeFiFusion is a dual-modal detector for price manipulation attacks (PMAs) in decentralized finance that fuses hand-crafted, manipulation-aware transaction-event features (basic, risk, and global families combined by fixed weights) with embeddings of smart-contract source code produced by a frozen, quantized DeepSeek-Coder-7B-Instruct-v1.5 backbone; fused sequences are classified by a DPF-Transformer with T5-style relative positional encoding. The authors assemble 225 verified PMA incidents from 2020-2026 ($804.3M total losses) plus 1,150 benign transactions. Under chronological five-fold cross-validation, DeFiFusion recalls 222 of 225 attacks (98.67%) at 96.10% precision, against 127, 117, and 120 recalls for DeFiRanger, DeFort, and DeFiScope — the first two re-implemented with LLM assistance, the last re-run with a substituted backbone. A simulated zero-day split (training through 2024, validation on 2025, testing on 18 incidents from 2026) achieves 100% recall with millisecond-scale online inference after offline contract pre-computation. Ablations show complementary modalities (disjoint miss sets) and smoother optimization with T5-RPE. Main limitations: a small positive class; benign data filtered from 428,523 transactions down to 1,150; partly re-implemented baselines; no isolated ablation for some heuristic signals such as digit-'9' counting; per-variant decision thresholds; no confidence intervals; and no explicit code URL.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- PMA maliciousness emerges only from the interaction between transaction behaviors and the exploited contract logic, so transaction-only or contract-only detection is fundamentally limited.
- DeFiFusion recalls 222 of 225 verified attacks (98.67%) at 96.10% precision under chronological five-fold cross-validation, versus 127, 117, and 120 for DeFiRanger, DeFort, and DeFiScope.
- In a simulated zero-day temporal split (train <=2024, test on 18 incidents from 2026), DeFiFusion achieves 100% recall with millisecond-scale online inference.
- T5-style relative positional encoding yields smoother optimization and better generalization than absolute positional encoding for cyclic execution patterns.
- The paper's 225-incident dataset (2020-2026, $804.3M losses, six chains) is the largest verified public collection of real-world PMA incidents.
Strengths
Largest verified PMA benchmark with per-case transparency
Table II enumerates all 225 incidents (2020-2026, six chains, $804.3M total) with per-tool verdicts; the >=$1M subset (52 cases, $781.68M) is totaled separately, and contract-side comparisons (DeFiTainter 34/95, SMARTCAT 83/194) are reported off-table.
Temporal zero-day evaluation design
RQ2 trains on 2020-2024, validates on 2025, tests on 18 authenticated 2026 incidents (Table III), avoiding retrospective leakage; per-case precompute (0.26-7.23 s) and online runtime (~4.6-22.3 ms total) are reported per incident.
Complementary-modality evidence supports the core thesis
Transparent, label-free feature engineering
Conservative FN accounting and honest failure analysis
Weaknesses
Curated benign set undermines precision as a deployment estimate
Baseline comparability
Post-hoc features conflict with the real-time claim
Unvalidated heuristic signals inside the risk score
Thin statistical reporting
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
DeFiFusion (arXiv:2609.11008) is a method paper on detecting Price Manipulation Attacks (PMAs) in DeFi. The authors argue PMAs are recognizable only through the interaction of transaction behavior and exploited contract logic, instantiated in three phases: manipulation-aware event encoding (~80 hand-crafted features per event), LLM-based contract semantic extraction (frozen DeepSeek-Coder-7B-Instruct-v1.5, 4-bit NF4 quantized, masked average pooling), and a DPF-Transformer classifier with T5-style relative positional encoding (T5-RPE). Headline claims: 222/225 recall (98.67%) at 96.10% precision under chronological five-fold CV; 100% recall on 18 real 2026 incidents in a temporal zero-day split; and the largest verified public PMA dataset (225 incidents, 2020-2026, $804.3M across six chains).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is DeFiFusion and what problem does it solve?
DeFiFusion is a detection framework for price manipulation attacks in decentralized finance (DeFi). It jointly models on-chain transaction events and smart contract source-code semantics, aiming to fix two failure modes: transaction-only detectors misclassifying legitimate volatility, and static contract analyzers flagging vulnerabilities that can never be exploited in practice.
How accurate is DeFiFusion?
On 225 verified attacks from 2020-2026 with chronological five-fold cross-validation, the paper reports 98.67% recall (222 of 225) and 96.10% precision against 1,150 benign transactions. In a simulated zero-day split testing on 18 real 2026 incidents, it reports 100% recall, with online inference taking roughly 5-22 milliseconds after offline contract pre-computation.
What datasets does the paper introduce?
A set of 225 verified price manipulation attacks across six blockchains (Ethereum, BNB Chain, Arbitrum, Polygon, Avalanche, Base) totaling $804.3M in losses, plus 1,150 benign transactions filtered from 428,523 by event-sequence length. The authors claim it is the largest verified public collection of real PMA incidents and state it is released.
Can DeFiFusion stop attacks in real time?
The authors argue yes: contract embeddings are precomputed offline (about 0.3-7.2 seconds per contract) and online event extraction plus classification totals roughly 5-22 ms, within block-confirmation windows. However, its features are defined over a complete incident's event log, so triggering detection mid-attack on partial sequences is not actually evaluated in the paper.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure-internal content is conveyed only via captions in the extracted text (scatter separations in Fig. 2, significance brackets in Figs. 4-5, loss-surface geometry in Fig. 3); judgments about those visuals rely on the authors' descriptions rather than the plotted values.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.