Adversarial Malware Generation in Linux ELF Binaries via Semantic-Preserving Transformations

Technical Summary: Adversarial Malware Generation in Linux ELF Binaries

1. Problem Statement

While significant research exists regarding adversarial attacks on Windows Portable Executable (PE) files, there is a notable research gap concerning the Linux Executable and Linkable Format (ELF). As Linux usage grows in cloud services, high-performance computing, and IoT, the vulnerability of machine learning (ML)-based malware detectors to ELF-based adversarial attacks becomes a critical security concern.

The core challenge is to generate adversarial malware: modified binaries that retain their original malicious functionality (semantic preservation) but are specifically engineered to evade detection by ML classifiers, causing them to be misclassified as benign.

2. Methodology

The authors propose a novel adversarial malware generator for the ELF format based on a genetic algorithm-inspired workflow. The methodology is structured around three pillars:

• Modification Types (12 total): The generator employs various techniques to perturb the binary without breaking its execution. These are categorized by risk level:
  • Low Risk (Structural): Adding/removing sections, modifying padding between loadable segments, rotating segments, or appending benign content.
  • Medium Risk (Metadata/Symbols): Modifying the .strtab section (adding benign symbols/removing malware symbols) or modifying non-executable section content.
  • High Risk (Code-level): Changing registers or interchanging instructions with equivalent logical properties.
• Data Sources (7 total): To populate these modifications, the generator uses diverse sources, including sequences of binary zeros, random data, static files (manual pages/text sections), extracted strings from benign files, and extracted symbols.
• Generator Workflow: The process is iterative. For each input malware file, the generator applies a combination of "iterative" modifications (which can be repeated) and "one-time" modifications. It uses a selection process to keep only the samples that result in the lowest malware confidence score from the target classifier.

Target Classifier: The researchers used MalConv, a widely recognized CNN-based deep learning model, as the target for evasion.

3. Key Contributions

• New Adversarial Generator: A specialized tool designed specifically for the ELF format using a diverse set of 12 modification types and 7 data sources.
• Enhanced Evaluation Metrics: Beyond the standard Evasion Rate (ER), the authors introduced:
  • Extended Evasion Rate (EER): A metric that accounts for the distribution of confidence scores across a threshold, providing a more granular view of how "close" a file is to being misclassified.
  • Mean Difference in Confidence (MD): A metric to measure the average reduction in the detector's confidence in the malware label.
• Interpretability via Logging: The generator produces detailed .log and .data files, allowing researchers to trace exactly which operations and data offsets were used to achieve evasion, addressing the "black-box" nature of ML attacks.

4. Results

The experimental results demonstrate the high effectiveness of the generator:

• Overall Performance: Using the full suite of modifications and data sources, the generator achieved an Evasion Rate (ER) of 67.74% and a Mean Difference in Confidence (MD) of -0.50.
• Most Effective Techniques: The analysis revealed that the most successful modifications were extending padding between loadable segments, modifying the .strtab section, adding new sections, and appending benign content.
• Data Source Sensitivity: The most successful data source was Extracted Strings. The results suggest that the MalConv classifier is highly sensitive to the presence of benign strings, regardless of their location within the ELF file.
• Risk vs. Reward: Interestingly, high-risk modifications (like changing instructions/registers) did not yield better results than low-risk structural modifications, suggesting that metadata and string manipulation are more efficient for evasion.

5. Significance

This research highlights a significant vulnerability in modern Linux security infrastructures. It demonstrates that ML-based malware detectors can be effectively bypassed using relatively "safe" structural modifications that do not alter the program's logic. By providing a framework for both attack (the generator) and evaluation (the new metrics), the paper provides a foundation for developing more robust, "adversarially trained" defense mechanisms for Linux environments.