Linux ELF Malware Generator Evades ML Detection With Semantic-Preserving Changes

As Linux continues to dominate high-performance computing, cloud services, and Internet of Things (IoT) devices, it has become a prime target for cybercriminals.

However, while much research has focused on manipulating Windows executables to bypass security, the Linux Executable and Linkable Format (ELF) has largely been ignored.

To address this gap, researchers at the Czech Technical University in Prague developed a new tool to assess how easily Linux malware can evade modern machine learning (ML) defenses.

Semantic-Preserving Transformations

Creating adversarial malware is a delicate balancing act. Attackers must modify the binary code to confuse antivirus software while ensuring the malicious program still works perfectly.

The researchers achieved this by using “semantic-preserving transformations”, modifications that change the file’s underlying signature to security scanners without breaking its original execution flow.

The newly designed generator relies on a simplified genetic algorithm to process these changes.

It automatically explores the modification space using 12 different types of alterations and 7 distinct data sources. Some of the key modifications include:

• Adding new sections near the end of the ELF file.
• Modifying the unused padding space between loadable segments.
• Appending benign file content to the end of the malware executable.
• Modifying the static symbols within the .strtab string table section.

Tricking Machine Learning Detectors

To test their generator, the researchers targeted MalConv, a widely used ML-based malware detector. The evasion tool proved to be highly successful.

When all modification types and data sources were enabled, the generator achieved an Evasion Rate (ER) of 67.74%. This means more than two-thirds of the modified malware successfully slipped past the AI-powered defence system.

Furthermore, the modifications reduced the detector’s confidence in its malware classification by an average of -0.50.

A fascinating discovery from the study was how easily the ML model was tricked by standard text. By analyzing the generator’s advanced logs, the researchers found that simply injecting typical strings from benign files was incredibly effective.

The most successful evasion techniques involved modifying the .strtab section, adding symbols typical of clean files to the malware.

This reveals a significant structural weakness: the target ML classifier appeared highly sensitive to benign strings placed anywhere within the executable file.

This research highlights a growing challenge in the cybersecurity landscape. While machine learning is heavily relied upon to detect unknown threats, this study proves that ML models can be fooled with simple, carefully crafted tweaks.

As Linux usage continues to expand globally, security vendors must update their scanning solutions to look beyond simple string analysis.

Follow us on Google News, LinkedIn, and X to Get Instant Updates and Set GBH as a Preferred Source in Google.