Research map / Label-efficient and robust learning / Knowledge distillation
MTVHunter: Smart Contracts Vulnerability Detection Based on Multi-Teacher Knowledge Translation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MTVHunter, a multi-teacher knowledge distillation framework that detects four smart contract vulnerability types from bytecode by combining an abstract-pattern instruction denoising teacher and a source-code semantic teacher trained with neuron distillation.
Problem
Detecting smart contract vulnerabilities from bytecode is hard because vulnerability-irrelevant instructions add noise and bytecode lacks complete control- and data-flow semantics present in source code.
Approach
Two teachers guide a two-stage student: an instruction denoising teacher uses abstract vulnerability patterns (state-loading, state-operation, state-storage) to score CFG nodes, and a semantic complementary teacher maps source-code graph features to missing bytecode semantics via neuron distillation, cast as a regression task.
Key results
- MTVHunter reaches 95.10% accuracy and 93.30% F1 on Reentrancy, versus 70.39% accuracy for the best traditional tool (Slither) and 86.86% for the best GNN baseline (DMT).
- On TX.Origin, MTVHunter reports 98.39% accuracy and 96.77% F1, improving over DMT's 91.12% accuracy by about 7 points.
- Neuron distillation outperforms Vanilla-KD, CTKD, OFA-KD and FT on Reentrancy accuracy (95.10% vs 50.58%, 60.58%, 83.72%, 88.03%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add IDT with complete denoising (ablation) | MTVHunter · Reentrancy vulnerability detection | accuracy: 72.55 -> 90.98 | Helped |
| use partial rather than complete denoising (ablation) | MTVHunter · TX.Origin vulnerability detection | accuracy: 73.39 -> 85.48 | Helped |
| use partial rather than complete denoising (ablation) | MTVHunter · TX.Origin vulnerability detection | F1: 83.08 -> 72.73 | Hurt |
| add SCT (ablation) | MTVHunter · Reentrancy vulnerability detection | accuracy: 84.12 -> 95.10 | Helped |
| increase neuron count from 1 to 3 (ablation) | MTVHunter neuron distillation · Reentrancy vulnerability detection | F1: 86.62 -> 93.30 | Helped |
| increase neuron count from 2 to 3 (ablation) | MTVHunter neuron distillation · Timestamp vulnerability detection | F1: 85.52 -> 92.72 | Helped |
| use 4 rather than 3 neurons (ablation) | MTVHunter neuron distillation · smart contract vulnerability detection | performance | No clear effect |
| scale α and β by the same factor while keeping their ratio equal (ablation) | MTVHunter multi-knowledge loss · TX.Origin vulnerability detection | accuracy: 98.39% with 0.01: 1 and 0.02:2 | No clear effect |
| use neuron distillation instead of Vanilla-KD (ablation) | MTVHunter · Reentrancy vulnerability detection | accuracy: 50.58 -> 95.10 | Helped |
| use neuron distillation instead of CTKD (ablation) | MTVHunter · Reentrancy vulnerability detection | accuracy: 60.58 -> 95.10 | Helped |
| use neuron distillation instead of OFAKD (ablation) | MTVHunter · Reentrancy vulnerability detection | accuracy: +11.38% | Helped |
| use neuron distillation instead of FT (ablation) | MTVHunter · Reentrancy vulnerability detection | accuracy: +7.07% | Helped |
Surprising findings
- Partial denoising in the instruction denoising teacher sometimes performs worse than no denoising, with the authors arguing that it confuses noise and vulnerability features.
- Using 2 neurons is not uniformly better than 1 neuron; on Timestamp and TX.Origin, 1-neuron gives higher accuracy and precision while 2-neuron has higher recall and F1.
Limitations
- Labels for the 10,356 contracts were produced by manual auditing and AVP-based classification, and the dataset is not stated as released.
- Only four vulnerability types are covered, and the relevant chains use a maximum path length of l<6.
- Reported gains rely on a single split with small positive counts for some categories (e.g., 15 nodes in case study and few positives per vulnerability type).
Experimental setup
Datasets: 229,178 public Ethereum smart contracts (filtered to 10,356 compiled contracts)
Benchmarks: in-house smart contract vulnerability dataset (Reentrancy, Timestamp, TX.Origin, Delegatecall)
Metrics: accuracy recall precision F1
Baselines: SmartCheck Mythril ContractFuzzer Oyente Securify Slither Vanilla-RNN Bi-LSTM GAT TMP SMS AME
Builds on: Vanilla-KD CTKD OFA-KD DMT AME
Closest papers by meaning
- Detecting Bugs with Substantial Monetary Consequences by LLM and Rule-based Reasoning (NeurIPS 2024 · similarity 0.86)
- SCDBench: A Benchmark for LLM-Based Smart Contract Decompilers (NeurIPS 2026 · similarity 0.85)
- CLEP: A Novel Contrastive Learning Method for Evolutionary Reentrancy Vulnerability Detection (AAAI 2025 · similarity 0.85)
- EVMbench: Evaluating AI Agents on Smart Contract Security (ICML 2026 · similarity 0.84)
- SCALM: Detecting Bad Practices in Smart Contracts Through LLMs (AAAI 2025 · similarity 0.84)
- BlockScan: Detecting Anomalies in Blockchain Transactions (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.