SCOUT: Smart Contract Open-Set Unknown Vulnerability Detection via Multi-Perspective Feature Fusion and Anomaly-Aware Recognition
Year: 2026
Type: article
Abstract: Smart contracts secure tens of billions of dollars in digital assets, and their immutability makes them a natural target for attackers. Current detection methods lean on predefined rules or closed-set classifiers, which limits their coverage of previously unseen vulnerabilities. We propose SCOUT (Smart Contract Open-set Unknown-vulnerability Detector), a framework that couples multi-perspective feature fusion with anomaly-aware open-set detection. SCOUT instruments the Ethereum Virtual Machine (EVM) to collect opcode sequences, call depths, and runtime parameters; symbolic simulation then maps numerical parameters to 22 semantic types while preserving fine-grained data-flow information. Hybrid embeddings encode opcode semantics, a graph convolutional network models data dependencies over a sparse dependency graph, and four statistical metrics (jump flags, gas costs, call-depth changes, and storage access flags) complete the fused representation. An anomaly-driven rule then combines a learned anomaly score with per-class prototype-distance thresholds to identify unknown vulnerabilities. From a symmetry perspective, SCOUT learns class-symmetric representations while preserving inter-class asymmetry and detects unknown vulnerabilities as symmetry-breaking samples. Six-fold leave-one-type-out cross-validation on 10,508 opcode traces yields a mean unknown-detection rate of 52.7%. When Unchecked External Call is held out, SCOUT reaches 90.2%, exceeding classical out-of-distribution baselines.
Source: Symmetry
DOI: 10.3390/sym18111919
Authors: Peiqiang Li, Yin Gu, Xiaofei Xing
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
Awards: G6404655336, G9001376548
OpenAlex ID: W7234055213
