[Paper] StriderSPD: Structure-Guided Joint Representation Learning for Binary Security Patch Detection

Source: arXiv - 2601.05772v1

Overview

The paper introduces StriderSPD, a novel framework that detects security patches directly from binary executables—without needing source code. By combining structural graph information with a large language model (LLM), the authors achieve more reliable binary‑level vulnerability‑fix identification, a capability that is especially valuable for closed‑source software.

Key Contributions

• Structure‑guided joint representation: Integrates a graph‑based branch (capturing control‑flow and data‑flow structures) with an LLM that processes assembly/pseudo‑code tokens.
• Adapter design for token‑level alignment: Custom adapters map graph embeddings onto the LLM’s token space, enabling seamless fusion of structural and textual cues.
• Two‑stage training strategy: Addresses the imbalance between the massive LLM and the lightweight graph branch, ensuring both components learn effectively.
• Realistic, disjoint benchmark: Constructs a new binary security‑patch detection dataset that is project‑ and domain‑disjoint from existing corpora, providing a more faithful evaluation of closed‑source scenarios.
• Empirical superiority: Demonstrates that StriderSPD outperforms prior binary SPD methods across precision, recall, and F1‑score on the new benchmark.

Methodology

1. Input preprocessing

   • Binary files are disassembled into assembly code and transformed into pseudo‑code (a higher‑level, language‑like representation).
   • From these, a program graph is built, encoding control‑flow (CFG) and data‑flow relationships between instructions.

2. Dual‑branch architecture

   • LLM branch: A pretrained large language model (e.g., CodeBERT or similar) consumes the tokenized assembly/pseudo‑code, learning textual semantics.
   • Graph branch: A lightweight Graph Neural Network (GNN) processes the program graph, extracting structural embeddings for each instruction node.

3. Adapter layers

   • Small trainable adapters sit between the GNN and the LLM, projecting graph embeddings onto the LLM’s token dimension.
   • This alignment lets the model attend to both syntax (tokens) and semantics (graph context) simultaneously.

4. Two‑stage training

   • Stage 1: Freeze the LLM, train the graph branch + adapters to learn a good structural mapping.
   • Stage 2: Unfreeze the LLM and fine‑tune the whole system jointly, using a balanced loss that prevents the massive LLM from dominating optimization.

5. Detection head

   • A classification layer predicts whether a given binary fragment corresponds to a security patch (binary label).

Results & Findings

• Ablation studies show that removing the graph branch drops F1 by ~7 points, confirming the value of structural cues.
• The two‑stage training yields a ~4% F1 gain over naïve end‑to‑end training, highlighting the importance of handling parameter imbalance.
• On the disjoint benchmark, StriderSPD maintains high performance, indicating strong generalization across unseen projects and domains.

Practical Implications

• Closed‑source security monitoring: Vendors and security teams can automatically flag patched binaries in the wild, even when vendors do not publish changelogs.
• Supply‑chain risk assessment: Integrate StriderSPD into CI/CD pipelines to verify that third‑party binaries have received the latest security fixes.
• Vulnerability management tools: Enhance existing scanners (e.g., CVE trackers) with binary‑level patch detection, reducing reliance on source‑code repositories.
• Incident response: Quickly determine whether a compromised binary has been updated, informing remediation timelines.

Limitations & Future Work

• Dependency on high‑quality disassembly: Obfuscated or heavily optimized binaries may produce noisy assembly/pseudo‑code, degrading detection accuracy.
• Scalability of graph construction: Large binaries can generate massive graphs; future work could explore hierarchical or sampling‑based GNNs.
• Model size: The LLM component remains heavyweight, which may limit deployment on edge devices; distillation or lightweight alternatives are worth investigating.
• Broader patch semantics: Current work focuses on binary patches that fix vulnerabilities; extending to feature‑level or performance patches is an open direction.

Authors

• Qingyuan Li
• Chenchen Yu
• Chuanyi Li
• Xin-Cheng Wen
• Cheryl Lee
• Cuiyun Gao
• Bin Luo

Paper Information

• arXiv ID: 2601.05772v1
• Categories: cs.SE, cs.CR
• Published: January 9, 2026
• PDF: Download PDF