TL;DR
This paper introduces MiniMal, a novel black-box adversarial attack method for static malware detectors. MiniMal significantly reduces the perturbation rate of adversarial malware samples while ensuring their functional integrity, achieving over 98% attack success against leading ML detectors and outperforming existing methods.
What Problem Does This Paper Address?
Static malware detectors, crucial for identifying malicious software, are increasingly relying on machine learning (ML) techniques. While these ML-based detectors offer improved generalization capabilities, they are vulnerable to adversarial attacks. Existing adversarial attack methods, particularly in black-box, hard-label scenarios (where only the final prediction is known), often suffer from high perturbation rates, rendering the adversarial samples unusable or unstable. Furthermore, these attacks frequently compromise the functional integrity of the malware, meaning the modified malware might not execute its intended malicious actions. This paper addresses the critical need for adversarial attacks that are both effective in evading detection and minimally disruptive to the malware's core functionality.
Methods
MiniMal employs a multi-stage approach to generate effective and minimally perturbed adversarial malware samples:
- Initialization of Adversarial Examples: The process begins by iteratively selecting perturbation actions and content using a roulette wheel selection algorithm. This algorithm leverages pre-computed weights based on the frequency of actions and content used in previously successful attacks, increasing the likelihood of selecting effective strategies.
- Optimization of Adversarial Examples: This stage focuses on minimizing perturbations.
- Action Minimization: A decremental loop mechanism is used to identify and remove redundant perturbation actions. By systematically removing actions and checking if the adversarial example still evades detection, the method prunes less critical modification steps.
- Content Optimization: * Binary Deletion: For scenarios with a single remaining action, binary deletion is applied to the perturbation content (treated as a byte array). This method efficiently searches for and removes non-critical bytes by recursively dividing the array.
- Particle Swarm Optimization (PSO) Deletion: When multiple actions remain, PSO is employed to optimize the perturbation content. PSO treats the problem as a search in a multi-dimensional space, where each particle represents a potential solution for distributing perturbation across different actions. A fitness function balances attack success, perturbation size, and functional integrity.
- Functionality Verification: Throughout the generation process, a dedicated module ensures the functional integrity of the adversarial examples. This involves parsing file format features and comparing control flow graphs (CFGs) of the original and modified malware. A high similarity in CFGs and preserved file structure indicates that the malware's core functionality remains intact.
Datasets
The study utilizes publicly available datasets for evaluation:
- Malware Detection PE-Based Dataset: Sourced from previous work, this dataset includes five malware types (Locker, Mediyes, Winwebsec, Zbot, Zeroaccess) and 1,000 benign software samples.
- MalwareBazaar: Additional malware samples (Trojan, Backdoor, Ransomware) were downloaded from this source.
- Selected Samples: For compatibility with target detectors, the authors selected 2,642 malware samples and 473 benign samples that were correctly classified by the detectors.
Key Insights and Ideas
- The Power of Minimal Perturbation: The core insight is that effective adversarial attacks don't require drastic modifications. By carefully selecting and optimizing perturbations, MiniMal achieves high evasion rates while keeping the malware functional and its size increase minimal.
- Functionality is Key for Real-World Attacks: The paper strongly emphasizes that for adversarial malware to be practically useful (e.g., for attackers), it must retain its malicious behavior. Their integrated functionality verification is a crucial differentiator.
- Roulette Wheel Selection for Efficient Initialization: Using a weighted selection mechanism based on past successes significantly improves the initial generation of adversarial examples compared to random selection.
- PSO for Complex Optimization: PSO proves effective in optimizing perturbation content when multiple modification actions are involved, handling the combinatorial nature of the problem.
Key Results
- High Attack Success Rate: MiniMal achieved attack success rates of 99.96%, 98.56%, and 99.62% against three leading ML-based malware detectors (MalConv, Ember, MalGCG), outperforming existing methods by 4.8% to 7.1%.
- Drastically Reduced Perturbation Rate: The average perturbation rates were consistently below 40% (38.91%, 32.18%, 33.92%), which is 9 to 11 times lower than previous methods. Notably, MiniMal achieved over 80% success even under a strict 100% perturbation limit.
- 100% Functional Integrity: Experimental verification using Cuckoo Sandbox confirmed that adversarial examples generated by MiniMal retained 100% functional integrity.
- Real-World Antivirus Performance: Against ClamAV, MiniMal achieved a 79% success rate, outperforming baseline methods. Against Avast, it achieved a 35% success rate, which is still competitive.
Limitations and Future Work
- Avast Performance: The paper notes that MiniMal performed poorly against Avast. The authors suspect this is due to Avast's incorporation of dynamic behavioral analysis, which their static modification approach may not effectively bypass.
- Future Work: The authors suggest exploring dynamic behavioral analysis as a potential avenue for future research to address detectors that rely on such mechanisms. Further investigation into optimizing the functionality verification process could also be beneficial.
Summary updated