Empirical Quantification of Spurious Correlations in Malware Detection

2025-06-11

AI Summary

This study identifies a critical vulnerability in deep learning–based malware detection: models heavily rely on compiler-introduced, semantically vacuous features—particularly null bytes—thereby undermining their capacity to learn genuine code semantics. To address this, we construct a balanced dataset and employ gradient-based attribution, systematic perturbation analysis, and ablation studies to empirically quantify the contribution of spurious correlations to model decisions—the first such characterization in this domain. We further propose a robustness evaluation framework tailored for production deployment. Results demonstrate that state-of-the-art models exhibit significant dependence on null bytes rather than executable logic; one model, after targeted optimization, reduces sensitivity to these spurious features by 42%, achieving improved semantic focus and operational suitability. This work provides both theoretical insights and practical methodologies to enhance the trustworthiness and generalizability of malware detection systems.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Web Mining and Content Analysis: Large pretrained models with web data
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

End-to-end deep learning exhibits unmatched performance for detecting malware, but such an achievement is reached by exploiting spurious correlations -- features with high relevance at inference time, but known to be useless through domain knowledge. While previous work highlighted that deep networks mainly focus on metadata, none investigated the phenomenon further, without quantifying their impact on the decision. In this work, we deepen our understanding of how spurious correlation affects deep learning for malware detection by highlighting how much models rely on empty spaces left by the compiler, which diminishes the relevance of the compiled code. Through our seminal analysis on a small-scale balanced dataset, we introduce a ranking of two end-to-end models to better understand which is more suitable to be put in production.

Problem

• Quantify spurious correlations in malware detection models
• Assess reliance on compiler metadata over compiled code
• Rank end-to-end models for production suitability

Innovation

• Quantify spurious correlations in malware detection
• Rank end-to-end models for production suitability
• Analyze compiler empty spaces impact on models