💻 computer science

CITADEL: A Semi-Supervised Active Learning Framework for Malware Detection Under Continuous Distribution Drift

CITADEL is a semi-supervised active learning framework that addresses concept drift in Android malware detection by introducing malware-specific feature augmentations and a multi-criteria selection strategy, achieving superior accuracy and efficiency on large-scale benchmarks with only 40% labeled data.

Original authors: Md Ahsanul Haque, Md Mahmuduzzaman Kamol, Suresh Kumar Amalapuram, Vladik Kreinovich, Mohammad Saidur Rahman

Published 2026-02-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Md Ahsanul Haque, Md Mahmuduzzaman Kamol, Suresh Kumar Amalapuram, Vladik Kreinovich, Mohammad Saidur Rahman

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a security guard at a massive, ever-changing city gate. Your job is to stop bad guys (malware) from entering while letting good citizens (benign apps) pass through.

In the past, you had a "Wanted Poster" with a photo of every criminal. But here's the problem: The criminals are shape-shifters. Every month, 300,000 new criminals show up, and they change their clothes, their faces, and even their names. If you only rely on your old Wanted Posters, you'll start letting the bad guys in because they don't look like the photos anymore. This is what experts call Concept Drift.

Furthermore, hiring a human expert to look at every single new person and decide if they are a criminal is impossible. There are too many people, and experts are slow.

Enter CITADEL. Think of it as a super-smart, self-learning security system designed to handle this chaos. Here is how it works, broken down into simple analogies:

1. The Problem: The "Shape-Shifting" Criminals

Traditional security systems are like a teacher who only learns from a textbook written in 2010. When the world changes in 2024, the teacher gets confused.

The Issue: Malware changes its behavior constantly. A virus that used to steal SMS messages might now steal data via encrypted web traffic. To a computer, these look like completely different things, even though the "intent" is the same.
The Old Way: Experts tried to retrain the system constantly, but they needed a human to label every single new sample. With 300,000 new samples a month, this is like trying to drink from a firehose.

2. The Solution: The "Guessing Game" (Semi-Supervised Learning)

CITADEL changes the game. Instead of waiting for a human to label every new person, it uses a Semi-Supervised approach.

The Analogy: Imagine you have a small group of known criminals (labeled data) and a huge crowd of unknown people (unlabeled data).
The Strategy: CITADEL looks at the unknown crowd and makes an educated guess: "This person looks 95% like a criminal." If the system is very confident, it treats that guess as a fact and uses it to teach itself. It learns from the "unknowns" without needing a human to check every single one.

3. The Secret Sauce: "Digital Makeup" (Feature Augmentation)

This is where CITADEL gets clever. Standard AI training often uses "augmentations" (like flipping an image upside down) to help the AI learn. But malware isn't a picture; it's a list of binary switches (0s and 1s) representing things like "Does this app ask for SMS permission?"

You can't just flip a picture of a virus. So, CITADEL invented two new tricks to simulate how criminals change:

Bernoulli Bit Flip: Imagine the criminal suddenly decides to wear a hat instead of a mask. The system randomly flips a few switches in the data (e.g., "No longer asks for SMS, but now asks for Contacts"). This teaches the AI: "Even if the criminal changes a few details, they are still a criminal."
Bernoulli Feature Mask: Imagine the criminal hides a tool in their pocket. The system randomly "hides" (masks) some features in the data. This teaches the AI: "Don't panic if a feature is missing; look at the bigger picture."

By practicing on these "fake" variations, the AI becomes robust. It learns to recognize the essence of the malware, not just the specific details.

4. The Smart Filter: "Who Do We Ask?" (Active Learning)

Even with the guessing game, the system still needs to ask a human expert for help sometimes. But asking about everyone is too slow. CITADEL uses a Multi-Criteria Strategy to pick the best people to ask about.

Instead of asking random people, it looks for three specific types of "suspicious" individuals:

The Borderline Cases: People who look almost exactly like a criminal but might be innocent. (The system is unsure).
The Outliers: People who look nothing like anyone the system has ever seen before. (They are far away from the known groups).
The Low-Confidence Guesses: People the system is just guessing about with low certainty.

By only asking the human expert about these specific, tricky cases, CITADEL gets the most "bang for its buck." It learns faster with fewer questions.

5. The Result: A Faster, Smarter Guard

The paper tested CITADEL on four massive datasets spanning over a decade of Android history.

Performance: It beat the previous best systems by a significant margin (up to 14% better accuracy), even when it only had 40% of the data labeled by humans.
Speed: It is incredibly efficient. It trains 24 times faster and does 13 times fewer calculations than the previous top method.

The Big Picture

Think of CITADEL as a self-driving security guard that doesn't just memorize faces. Instead, it:

Imagines how criminals might disguise themselves (Augmentation).
Guesstimates the identity of the crowd (Semi-Supervised Learning).
Strategically asks a human boss only for the trickiest cases (Active Learning).

This allows it to stay ahead of the constantly evolving world of Android malware without needing an army of human experts to label every single new threat. It's a smarter, faster, and more scalable way to keep our digital cities safe.

1. Problem Statement

The paper addresses the critical challenge of concept drift in Android malware detection. Over time, the statistical properties of both malware and benign applications evolve (e.g., new evasion techniques, changing API usage), causing machine learning detectors to degrade in performance.

Key challenges identified include:

Scalability of Labeling: With ~300,000 new malware samples emerging monthly, expert labeling is a bottleneck. Human analysts can only review ~80 samples/day, leaving the vast majority of data unlabeled.
Limitations of Existing Methods:
- Fully Supervised Approaches: Rely on continuous access to labeled data, which is unrealistic for long-term drift.
- Existing Semi-Supervised Learning (SSL): Methods like FixMatch are designed for computer vision (using rotations/flips) and fail when applied directly to high-dimensional, structured binary malware features.
- Existing Active Learning (AL): Current AL strategies often rely on single uncertainty metrics (e.g., margin) and fail to adapt to abrupt, unpredictable drifts over long temporal spans (e.g., 12-year datasets).

2. Methodology: The CITADEL Framework

CITADEL is a Semi-Supervised Active Learning (SSAL) framework designed to adapt to long-term distribution shifts using limited labeled data. It integrates three core components:

A. Domain-Specific Feature Augmentations

Since standard image augmentations (rotation, color jitter) are invalid for binary malware feature vectors, CITADEL introduces two stochastic perturbations based on the Bernoulli distribution:

Bernoulli Bit Flip: Randomly flips bits in the binary feature vector (simulating the addition or removal of specific behaviors/APIs).
Bernoulli Feature Mask: Randomly zeros out a subset of features (simulating obfuscation or the disappearance of deprecated permissions).

Implementation: These are applied as Weak (low probability, e.g., $p=0.01$ ) and Strong (higher probability, e.g., $p=0.05$ ) augmentations within the FixMatch paradigm. Weak augmentations generate high-confidence pseudo-labels, while strong augmentations enforce consistency, regularizing the model against feature instability.

B. Enhanced Objective Function

CITADEL modifies the standard SSL loss to better handle boundary samples where benign and malware features overlap:

Supervised Contrastive Loss ( $L_{con}$ ): Added to the standard cross-entropy and unsupervised consistency losses. This loss pulls feature representations of the same class closer together and pushes different classes apart in the latent space, improving class separation for ambiguous samples.
Total Loss: $L = L_{sup} + \lambda_u L_{unsup} + \lambda_{con} L_{con}$ .

C. Multi-Criteria Active Learning (AL) Strategy

To select the most informative unlabeled samples for expert labeling under a constrained budget, CITADEL combines three complementary criteria:

Boundary Proximity (Softmax Margin): Selects samples where the difference between the top two predicted class probabilities is small (high uncertainty).
Feature Displacement ( $L_p$ -Norm Distance): Calculates the distance between an unlabeled sample and its nearest labeled neighbor in the embedding space. Large distances indicate semantic novelty or outliers.
Low Confidence: Selects samples where the maximum predicted probability is low.

Selection Mechanism: These three scores are normalized and summed to create a unified uncertainty score. The top- $k$ samples are selected for labeling, ensuring the model focuses on both boundary ambiguity and distributional outliers.

3. Key Contributions

CITADEL Framework: A novel SSAL framework tailored for Android malware detection under continuous concept drift, bridging the gap between SSL and domain-specific feature constraints.
Novel Augmentations: Introduction of Bernoulli Bit Flip and Feature Masking to effectively model feature-level variations in binary malware data, replacing ineffective image-based augmentations.
Multi-Criteria Selection: A robust sample selection strategy that jointly considers decision boundary margins, feature-space distances, and prediction confidence, outperforming single-metric AL approaches.
Supervised Contrastive Loss: Integration of contrastive learning to improve the separation of ambiguous boundary samples in the latent space.
New Benchmarking: Construction of a 12-year MaMaDroid-based dataset and extensive evaluation on four large-scale benchmarks (APIGraph, Chen-AZ, MaMaDroid, LAMDA).

4. Experimental Results

The framework was evaluated on four datasets spanning up to 12 years of data. Key findings include:

Performance Gains:
- APIGraph: F1 score of 93.5% (vs. 92.4% for Chen-AL).
- Chen-AZ: F1 score of 82.7% (vs. 75.5% for Chen-AL).
- MaMaDroid: F1 score of 44.9% (vs. 36.0% for Chen-AL).
- LAMDA (Most Drifted): F1 score of 77.7% (vs. 43.0% for Chen-AL), representing a ~14% improvement.
- Note: These results were achieved using only 40% labeled data.
Efficiency:
- CITADEL is 24× faster in training time compared to Chen-AL.
- It requires 13× fewer operations, making it highly scalable for real-time deployment.
Ablation Studies:
- Removing the multi-criteria selection or the contrastive loss resulted in significant performance drops.
- Standard FixMatch with image augmentations failed completely (0% F1) on malware data, validating the need for domain-specific augmentations.

5. Significance and Impact

Real-World Applicability: CITADEL addresses the "labeling bottleneck" by effectively leveraging abundant unlabeled data, making it viable for scenarios where expert labeling is scarce and expensive.
Long-Term Robustness: Unlike previous methods that degrade over time, CITADEL maintains high performance over 12-year spans, proving its ability to adapt to unpredictable, long-term concept drift.
Computational Efficiency: The drastic reduction in training time and operations allows for frequent model updates in response to emerging threats, a critical requirement for modern security operations.
Generalizability: The framework demonstrates that semi-supervised learning, when combined with domain-aware augmentations and multi-criteria selection, is a superior alternative to fully supervised approaches for evolving malware detection.

In conclusion, CITADEL represents a significant step forward in adaptive malware detection, offering a scalable, efficient, and robust solution to the continuous evolution of Android threats.