Abstract
The rapid growth of the Android application ecosystem presents a significant challenge for manual security vetting. This paper explores a semi-supervised learning approach using a hybrid label propagation technique to identify malware. By constructing a network of applications based on shared features and treating labeled and unlabeled data with distinct weights during propagation, the proposed method achieves superior performance compared to traditional hard and soft clamping techniques. Experimental results demonstrate high F1-measures even when labeled data is scarce or highly imbalanced.
Introduction
With millions of applications available on the Google Play Store, human-led manual vetting is unfeasible. While automated scanners provide some relief, the availability of "ground truth" labels—confirmed malware versus benign apps—is limited. This creates a classic semi-supervised learning problem: leveraging a small set of labeled data to classify a vast amount of unlabeled data. Standard approaches often struggle with noise or fail to maintain the integrity of original labels. This research addresses these challenges by modeling the problem as a network propagation task.
Methodology
The researchers developed a multi-step process to classify Android applications:
- Feature Extraction: Apps are converted into binary feature vectors consisting of 471 attributes, including requested permissions, intent actions, API calls, and obfuscation signatures.
- Network Construction: A graph is built where nodes represent apps and edges represent similarity. Edge weights are calculated using a Gaussian kernel based on the "closeness" of feature vectors.
- Hybrid Label Propagation: The core innovation is a hybrid variant of label propagation. Unlike Hard-clamping (fixed labels) or Soft-clamping (all labels can shift), the Hybrid Variant assigns a weight ($\alpha$) to each node. Labeled apps are assigned $\alpha = 0$ (absolute truth), while unlabeled apps are assigned $\alpha = 0.1$. This allows the network to maintain the integrity of ground truth while providing flexibility for unlabeled nodes to adapt based on their neighbors.
Datasets
The study utilized 20,000 Android apps (10,000 malicious and 10,000 benign). The models were tested on both balanced datasets and imbalanced datasets (1:100 ratio) to simulate real-world conditions where malware is a "needle in a haystack."
Results
The performance of the hybrid variant was evaluated using the F1-measure against standard clamping methods. The results across various data distributions are summarized below.
In balanced scenarios with only 50 labeled examples of each class, the Hybrid Variant achieved an F1-measure of 0.887, significantly higher than Hard-clamping (0.727) and Soft-clamping (0.728). As labeled data increased to 5,000 per class, the Hybrid Variant's performance scaled to an F1-measure of 0.979. In imbalanced scenarios (e.g., 50 labeled malware vs. 5,000 labeled benign), the Hybrid Variant maintained a lead with an F1 of 0.752.
Discussion and Conclusion
The primary advantage of the hybrid approach lies in its differentiated treatment of labeled and unlabeled data. By clamping ground truth more strongly, the model prevents the "washing out" of known labels by the larger volume of unlabeled data. The network topology effectively captures clusters of malicious behavior through structural similarities like shared API calls.
The model's robustness to data scarcity makes it particularly valuable for identifying emerging threats. Future work will focus on dynamic parameter tuning for $\alpha$, incorporating explicit mechanisms for class imbalance, and investigating the impact of noisy or incorrect ground truth labels on network accuracy.
Summary updated