A Deep Dive into Function Inlining and its Security Implications for ML-based Binary Analysis

This artifact accompanies the paper “A Deep Dive into Function Inlining and its Security Implications for ML-based Binary Analysis.”
It provides the complete codebase, datasets, and scripts necessary to reproduce the experiments and figures presented in the paper, submitted to NDSS 2026.

Contents

The artifact consists of three main .tar.gz archives. It is designed to examine the impact of extreme function inlining on machine learning–based security tasks, including binary similarity detection (T1), function name prediction (T2), malware detection (T3), malware family prediction (T4), and vulnerability detection (T5). The archives are organized as follows:

1. AE_NDSS_Code.tar.gz — Source code, analysis scripts, ML models, and utilities for dataset construction, feature extraction, and evaluation.
   Includes:

   • Dataset Construction i.e. compilation and inlining-sweep scripts

   • Modified TikNib feature extraction pipeline used mainly for T3 and T4

   • ML evaluation frameworks for all tasks 

   • Figure and results reproduction scripts

2. AE_NDSS_Code_Docker.tar.gz (Optional)— A prebuilt Docker image recommended if you only want to run the ML models and reproduce the main figures. It provides:

   • All Python/Conda dependencies

   • Preconfigured environments for T1/T2/T5

   • Runtime support for Docker-in-Docker (required for T1/T5 models)

   • System Python 3.9 (needed for T3/T4)

   • Scripts for generating the main figures

3. AE_NDSS_Dataset_VI.A.tar.gz — Provides the precompiled binary datasets used across evaluation tasks T1–T5, allowing reviewers to run experiments directly without rebuilding from source.
4. AE_NDSS_Dataset_VI.B.tar.gz — Contains the datasets compiled across varying compiler settings, systematically exploring configurations that approach extreme inlining behavior.
5. START_EVALUATION.md — Provides step-by-step setup, testing, and reproduction instructions for artifact reviewers.

Notes on Excluded Data

To comply with third-party licensing and redistribution restrictions, the following datasets are not included in this artifact:

• T4: The dataset originally curated from Alrawi et al. (“BadThings”) is available via their official repositories (Dropbox).

• T5: Firmware samples from Netgear and TP-Link are publicly referenced in prior work but are not redistributed here.

• SPEC CPU 2006 binaries: Not shared due to licensing restrictions. Scripts for reproducing related figures are included and runnable without these binaries.