Performance evaluation of deep neural network on malware detection: visual feature approach.

Area of Science:

• Computer Science
• Cybersecurity
• Artificial Intelligence

Background:

• Malware detection is crucial for securing devices against malicious applications.
• Deep neural networks are vulnerable to adversarial attacks, posing a significant threat.
• Understanding malware variant characteristics is key for effective detection.

Purpose of the Study:

• To assess the robustness of the DenseNet deep learning model against adversarial attacks in malware detection.
• To investigate the impact of different noise types, including Gaussian noise and Fast Gradient Sign Method (FGSM), on classifier performance.
• To evaluate DenseNet's effectiveness on distinct malware datasets (Malimg and BIG2015).

Main Methods:

• Utilized the DenseNet deep learning model for malware classification.
• Generated adversarial samples by adding Gaussian noise to malware datasets.
• Conducted experiments using the Fast Gradient Sign Method (FGSM) to create adversarial examples.

Main Results:

• DenseNet accurately identified modified malware samples with Gaussian noise on the Malimg dataset.
• A marginal decrease in classifier performance was observed on the BIG2015 dataset with Gaussian noise.
• Significant decreases in classification accuracy were detected for both datasets when using FGSM attacks.

Conclusions:

• Deep learning models, including DenseNet, require enhanced robustness against sophisticated adversarial attacks.
• Further research is needed to develop more resilient deep learning architectures for malware detection.
• Adversarial attack methods like FGSM pose a considerable challenge to current deep learning-based security systems.