Abstract
In the rapidly evolving landscape of cybersecurity, deep learning models offer high accuracy for malware detection but often lack transparency. This paper presents an approach using Logic Explained Networks (LENs) to provide human-readable First-Order Logic (FOL) explanations without sacrificing predictive performance. We introduce the "Tailored-LEN" method, which optimizes the trade-off between rule complexity and fidelity. Evaluation on the EMBER dataset demonstrates that LENs achieve approximately 92% accuracy, comparable to state-of-the-art black-box models, while offering actionable insights for security experts.
Introduction
Traditional signature-based detection is increasingly ineffective against the volume of modern "zero-day" threats. While Deep Neural Networks (DNNs) excel at identifying novel malware, their "black box" nature poses a significant hurdle in safety-critical environments where trust and accountability are paramount. Security administrators require clear reasoning before acting on a detection. Existing interpretable models, such as decision trees, often fail to handle the high-dimensional complexity of malware data, leading to a performance-explainability trade-off. This work addresses this gap by leveraging Logic Explained Networks to provide both high accuracy and clear logical justifications.
Methodology: Logic Explained Networks (LENs)
The core of our approach is the Logic Explained Network (LEN), a neural architecture designed for inherent explainability. LENs utilize human-understandable predicates (e.g., binary features like has_signature or is_dll) as inputs.
Local and Global Explanations
The system operates at two levels of granularity:
- Local Explanations: For a specific file, the model provides a logic formula explaining its classification (e.g.,
¬is_dll ∧ ¬has_debug). - Class-level (Global) Explanations: By aggregating local rules, the system derives broader logic formulas that characterize entire malware classes.
The Tailored-LEN Innovation
Standard LENs can produce overly complex rules or suffer from high false-positive rates on large datasets. We introduce "Tailored-LEN," which employs a line search optimization to determine the optimal precision threshold for rule combination. This method filters outliers and simplifies the resulting FOL formulas, ensuring they remain interpretable for human experts while maintaining high fidelity to the underlying model's behavior.
Experimental Setup
We evaluated our approach using the EMBER dataset, a benchmark containing 800,000 labeled Windows Portable Executable (PE) files. We utilized a version of EMBER with derived binary features (an ontology), enabling the model to process thousands of interpretable "true/false" characteristics.
Results and Discussion
Our experiments reveal that LENs effectively bridge the performance gap between interpretable and black-box models.
Performance Benchmarks
LENs achieved an accuracy and F1-score of approximately 92% on the EMBER dataset using over 100 features. This performance is within a 3-5% margin of top-tier black-box models like LightGBM. Crucially, as shown in the figure below, the LEN-ESFS variant consistently matches or outperforms other interpretable "concept learning" baselines across various feature counts.
Expert Validation and Fidelity
A key finding is the high fidelity of the generated rules. When presented to cybersecurity experts, the extracted logic formulas—such as those describing "packed" malware (e.g., has_section_high_entropy ∧ ¬is_dll ∧ ¬has_debug)—were confirmed to align with real-world malicious behaviors. The Tailored-LEN method successfully navigated the Pareto front of explanation, providing rules that are both simple enough for human verification and accurate enough for operational use.
Conclusion and Future Work
This work demonstrates that logical explainability does not have to come at the cost of detection accuracy in malware analysis. While feature binarization introduces a slight performance ceiling, the benefits of human-readable justifications are substantial for threat hunting. Future research will focus on scaling to larger feature sets, tracking the evolution of logic rules as malware tactics shift, and integrating these explanations into real-time security dashboards.
Summary updated