.png){kind=icon}
Return
Evolutionary Black-Box Optimization for Functionality-Preserving Adversarial Malware Generation in Consumer Systems
.png){kind=icon}
PRE
AI
DOI：10.1109/tce.2026.3734301.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Recent cyberattacks show that Windows systems remain prime targets, with adversaries increasingly using stealthy, behavior-preserving techniques that evade traditional detection. Given their central role in consumer electronics and consumer computing environments, this highlights the need for robust malware detection mechanisms capable of handling realistic, adversarial threats. In this work, we present a functionality-preserving black-box adversarial attack framework for Windows Portable Executable (PE) malware under practical constraints, including limited query budgets and executable validity. The proposed approach employs section injection and overlay padding to generate evasive samples without altering execution semantics and leverages the CMS-ES, a query-efficient optimization method that does not require access to model internals. The attack is evaluated on a balanced dataset of 30,000 samples and tested across five classifiers: MalConv, CNN, GBDT, Random Forest, and SVM, under query budgets of 10–80. Experimental results show that the proposed attack achieves up to 92% attack success rate (ASR) against the evaluated ML-based malware classifiers. In a separate VirusTotal-based evaluation, section-injected variants reduced the average number of detections from 58.20 to 38.41 engines. The generated samples also exhibit strong cross-model transferability and show improved query efficiency and perturbation control compared with mutation-based search under the evaluated settings. These findings expose critical weaknesses in machine learning–based malware detectors and emphasize the need for more robust, adversarially resilient defense strategies.
Keywords:
Consumer computing environment
Malware Classifiers
Adversarial Samples
Perturbations
Machine Learning Models
Transferability
PE Files
Journal
IF:
10.9
Papers:
5.3K
Citations:
6.8K
Organization
Cited Papers
No cited papers available