Effectiveness of Moving Target Defenses for Adversarial Attacks in ML-Based Malware Detection

Aqib Rashid, Jose SuchIEEE TDSC 2023 · OpenReview

TL;DR

This paper investigates the effectiveness of Moving Target Defenses (MTDs) against adversarial attacks in ML-based malware detection. The authors demonstrate that even sophisticated MTDs can be evaded with carefully crafted transferability and query-based attacks, and that attackers can potentially fingerprint defense mechanisms to gain an advantage.

What Problem Does This Paper Address?

Machine learning models are increasingly used for malware detection due to their ability to generalize and identify novel threats. However, these models are susceptible to adversarial attacks, where malicious samples are subtly modified to be misclassified as benign. Moving Target Defenses (MTDs) have emerged as a promising defense strategy, aiming to increase attacker difficulty by dynamically changing defense configurations. This paper addresses the critical question of whether these MTDs truly provide robust protection against adversarial attacks in the malware detection domain, and if so, which MTD designs are most effective.

Methods

The authors propose and evaluate novel attack strategies, alongside existing ones, to test MTDs under various threat models:

  • Transferability Attacks: These attacks leverage the property that adversarial examples generated for one model can often fool another. The paper introduces a novel strategy that maximizes adversarial example transferability across a diverse ensemble of substitute models before targeting the MTD.
  • Query Attacks: These attacks iteratively probe the target model to generate adversarial examples without needing a substitute model. The paper adapts existing query attack methods for the discrete feature space of malware detection and proposes a heuristically-driven gray-box query attack that prioritizes features based on their frequency in benign samples.
  • Fingerprinting and Reconnaissance: The paper explores methods for attackers to gain insights into the MTD's internal workings, such as identifying its predictive nature (static vs. dynamic) and potentially discovering key hyperparameters like query budgets.

Datasets

The study utilizes three publicly available datasets for malware detection:

  • DREBIN: An Android malware dataset with extracted static features.
  • SLEIPNIR: A Windows malware dataset with features derived from PE files.
  • AndroZoo: A large-scale Android dataset with VirusTotal reports.

The datasets were split into training, validation, and testing sets using an 80:20 ratio for training/testing, with further partitioning for validation.

Key Insights and Ideas

  • MTDs are not a silver bullet: Despite their dynamic nature, many evaluated MTDs are surprisingly vulnerable to well-crafted adversarial attacks, often performing no better than static defenses.
  • Attackers can adapt: Novel attack strategies, particularly those that maximize transferability across diverse models and leverage domain-specific heuristics in query attacks, significantly increase evasion rates.
  • Fingerprinting is feasible: Attackers can gain valuable intelligence about MTDs by observing their prediction patterns and identifying key operational parameters, which can then be used to tailor more effective attacks.
  • Diversity is key, but not sufficient: While MTDs that incorporate diverse constituent models and dynamic movement show better resilience, they are not immune to sophisticated attacks.
  • Static defenses are generally weaker: MTDs that exhibit static or predictable behavior are more easily targeted, highlighting the importance of true dynamism.

Key Results

  • High Evasion Rates: The proposed novel attack strategies achieved high evasion rates across datasets, with peak rates reaching up to 96.3% for DREBIN and 96.1% for AndroZoo in black-box transferability attacks. Query attacks also demonstrated significant success, with average evasion rates of 72% for DREBIN, 69% for SLEIPNIR, and 90.3% for AndroZoo.
  • Improved Attack Efficiency: The novel gray-box query attack strategy reduced the number of queries required for evasion compared to its black-box counterpart.
  • MTD Vulnerabilities: Defenses like DeepMTD and MTDeep performed poorly, exhibiting static behavior and being easily evaded. Even more robust MTDs like Morphence and StratDef could be bypassed.
  • Fingerprinting Success: The paper demonstrated that attackers can identify the predictive nature of defenses and, in the case of Morphence, estimate its query budget.
  • False Positive Rate Concerns: While some defenses offered good accuracy, the study highlighted that defenses like veto voting can have unacceptably high false positive rates, making them impractical for real-world malware detection.

Limitations and Future Work

  • Dataset Limitations: The study acknowledges the challenge of obtaining comprehensive and up-to-date malware datasets, which can limit the generalizability of findings.
  • MTD Configurations: The evaluation used configurations close to original papers; further research could explore the impact of different hyperparameter tuning.
  • Future Work Recommendations:
    • Avoid static MTDs: Focus on dynamic and hybrid MTDs with robust movement mechanisms.
  • Increase model diversity: Utilize a wider range of model families beyond just DNNs.
  • Independent movement: Ensure MTD movement is not solely reliant on user activity but incorporates independent mechanisms like time or randomness.
  • Hyperparameter security: Develop MTDs that are resilient to the leakage of hyperparameters.
  • System awareness: Integrate MTDs with greater system awareness to detect and react to attacks in progress, potentially through stateful detection and cyber-threat intelligence.

Summary updated

Abstract

Several moving target defenses (MTDs) to counter adversarial ML attacks have been proposed in recent years. MTDs claim to increase the difficulty for the attacker in conducting attacks by regularly changing certain elements of the defense, such as cycling through configurations. To examine these claims, we study for the first time the effectiveness of several recent MTDs for adversarial ML attacks applied to the malware detection domain. Under different threat models, we show that novel transferability and query attack strategies can increase the evasion rate by 50+% against these defenses across Android and Windows, with 90+% evasion rate in some cases. We also show that fingerprinting and reconnaissance are possible and demonstrate how attackers may obtain critical defense hyperparameters as well as information about how predictions are produced. Based on our findings, we present key recommendations for future work on the development of effective MTDs for adversarial attacks in ML-based malware detection.
BibTeX
@article{rashid2023effectiveness,
  author = {Aqib Rashid and Jose Such},
  title = {{Effectiveness of Moving Target Defenses for Adversarial Attacks in ML-Based Malware Detection}},
  journal = {IEEE Trans. Dependable Secur. Comput.},
  year = {2023},
  url = {https://openreview.net/forum?id=9w5HhsoFf7},
}
Figure