[Paper Review] Towards Interpretable Ensemble Learning for Image-based Malware Detection

Yuzhou Lin, Xiaolin Chang|arXiv (Cornell University)|Jan 13, 2021

Anomaly Detection Techniques and Applications39 references4 citations

TL;DR

This paper proposes an interpretable ensemble learning framework, IEMD, for image-based malware detection by combining a Selective Deep Ensemble Learning (SDEL) detector with an Ensemble Deep Taylor Decomposition (EDTD) method for pixel-level explanation. IEMD achieves 99.87% detection accuracy with high-fidelity, robust, and expressive explanations, demonstrating that interpretability and accuracy can coexist in malware detection systems.

ABSTRACT

Deep learning (DL) models for image-based malware detection have exhibited their capability in producing high prediction accuracy. But model interpretability is posing challenges to their widespread application in security and safety-critical application domains. This paper aims for designing an Interpretable Ensemble learning approach for image-based Malware Detection (IEMD). We first propose a Selective Deep Ensemble Learning-based (SDEL) detector and then design an Ensemble Deep Taylor Decomposition (EDTD) approach, which can give the pixel-level explanation to SDEL detector outputs. Furthermore, we develop formulas for calculating fidelity, robustness and expressiveness on pixel-level heatmaps in order to assess the quality of EDTD explanation. With EDTD explanation, we develop a novel Interpretable Dropout approach (IDrop), which establishes IEMD by training SDEL detector. Experiment results exhibit the better explanation of our EDTD than the previous explanation methods for image-based malware detection. Besides, experiment results indicate that IEMD achieves a higher detection accuracy up to 99.87% while exhibiting interpretability with high quality of prediction results. Moreover, experiment results indicate that IEMD interpretability increases with the increasing detection accuracy during the construction of IEMD. This consistency suggests that IDrop can mitigate the tradeoff between model interpretability and detection accuracy.

Motivation & Objective

• To address the lack of interpretability in deep learning models for image-based malware detection, especially in security-critical environments.
• To overcome the trade-off between model accuracy and interpretability in existing image-based malware detection systems.
• To develop a novel explanation method that provides reliable, high-quality pixel-level attribution for model predictions.
• To design a training strategy that enhances interpretability without sacrificing detection performance.

Proposed method

• Propose a Selective Deep Ensemble Learning (SDEL) detector that improves prediction accuracy by combining multiple deep neural networks selectively.
• Introduce Ensemble Deep Taylor Decomposition (EDTD), a method to generate pixel-level saliency maps explaining SDEL predictions.
• Develop mathematical formulas to evaluate explanation quality using fidelity, robustness, and expressiveness metrics on pixel-level heatmaps.
• Design a novel Interpretable Dropout (IDrop) technique to train the SDEL detector while enhancing explanation quality.
• Use EDTD to generate explanations for each prediction and validate them using quantitative metrics.
• Integrate SDEL, EDTD, and IDrop into a unified framework called IEMD for end-to-end interpretable malware detection.

Experimental results

Research questions

• RQ1 Can an ensemble deep learning model achieve high detection accuracy while maintaining high-quality interpretability in image-based malware detection?
• RQ2 How can pixel-level explanations be reliably generated and evaluated for deep learning models in malware detection?
• RQ3 Does improving model accuracy through ensemble learning also enhance the interpretability of predictions?
• RQ4 Can a training strategy like Interpretable Dropout (IDrop) effectively reduce the trade-off between accuracy and interpretability?
• RQ5 To what extent do fidelity, robustness, and expressiveness metrics correlate with detection performance in explanation quality?

Key findings

• IEMD achieves a detection accuracy of 99.87%, demonstrating state-of-the-art performance in image-based malware detection.
• EDTD produces higher-quality explanations than prior explanation methods, as measured by fidelity, robustness, and expressiveness metrics.
• The quality of interpretability, as measured by explanation metrics, increases alongside detection accuracy during model training.
• The proposed IDrop method successfully mitigates the traditional trade-off between model accuracy and interpretability.
• The correlation between detection accuracy and explanation quality suggests that higher accuracy models can naturally yield more reliable and interpretable predictions.