🤖 machine learning

LLM-Generated Samples for Android Malware Detection

This study demonstrates that while fine-tuned LLM-generated synthetic data can effectively augment scarce Android malware datasets without significantly compromising detection accuracy, it remains insufficient as a standalone training source compared to real-world data.

Original authors: Nik Rollinson, Nikolaos Polatidis

Published 2026-01-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nik Rollinson, Nikolaos Polatidis

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 trying to spot a specific type of thief in a crowded city. To do your job well, you need to study photos of real thieves so you recognize their faces, clothes, and habits. But what if there are only a few photos of a specific kind of thief, like "BankRobbers," and you need more to train your eye?

This paper explores a new way to get those extra photos: asking a super-smart AI (called a Large Language Model, or LLM) to draw fake photos of thieves based on the real ones it has seen. The researchers wanted to know: Can these AI-drawn photos help us catch real thieves, or do they just confuse the security guard?

Here is a simple breakdown of what they did and what they found:

The Setup: The "Photo Album" and the "AI Artist"

The researchers started with a real album of photos (a dataset called KronoDroid) containing pictures of three specific types of Android malware (bad apps):

BankBot: Thieves who steal banking info.
Locker/SLocker: Thieves who lock your phone screen.
Airpush/StopSMS: Thieves who spam ads or send secret texts.

They asked an AI artist (specifically a model called GPT-4.1-mini) to look at these real photos and draw new, fake ones that looked just like them. They tried to make the AI draw 50 to 150 fake photos for each type of thief.

The Three Experiments

To test if the AI's drawings were useful, they ran three different training scenarios for their security guards (machine learning models):

The "Real Only" Test: The guard studied only the real photos.
- Result: The guard became a master detective. They caught almost every thief with near-perfect accuracy. This is the gold standard.
The "Real + Fake" Test: The guard studied the real photos plus the AI-drawn fake photos.
- Result: The guard still did an amazing job, almost as good as the "Real Only" group. The fake photos didn't confuse them; they just added a little extra practice. It's like studying a real map and a few hand-drawn sketches of the same city; you still know the way.
The "Fake Only" Test: The guard studied only the AI-drawn fake photos and was then tested on real thieves.
- Result: This was a mixed bag.
  - For the BankRobbers, the guard was okay but missed about half the real thieves.
  - For the Phone Lockers, the guard was basically guessing (like flipping a coin).
  - For the Spam/Text Thieves, the guard did surprisingly well, catching most of them.
- Why the difference? The researchers found that the AI was better at drawing the "Spam/Text" thieves because they had more real examples to learn from and the AI practiced drawing them longer. The AI struggled to capture the complex details of the other two types.

The Big Takeaway

The paper concludes with a simple rule of thumb:

AI drawings are great for "filling in the blanks." If you don't have enough real photos of a specific thief, asking an AI to draw some fake ones can help your security guard get better without hurting their performance.
AI drawings are NOT a replacement. You cannot train a security guard only on AI drawings and expect them to catch real thieves. The fake photos miss subtle, real-world details that only exist in the real thing.

The Analogy of the "Cooking Class"

Think of it like learning to cook:

Real Data is tasting a real, perfectly cooked steak.
AI Data is the AI describing what a steak tastes like and writing a recipe for you.

If you taste the real steak and read the AI's recipe, you'll be a great cook. But if you only read the AI's recipe and never taste a real steak, you might make a dish that looks like a steak but doesn't taste like one. You might miss the secret ingredient that only exists in the real thing.

In short: Use AI to help you practice when you're short on real examples, but never rely on AI alone to do the real work.

Technical Summary: LLM-Generated Samples for Android Malware Detection

Problem Statement
Android malware continues to evolve rapidly through obfuscation and polymorphism, challenging both signature-based defenses and machine learning (ML) models. A critical bottleneck in developing robust detection systems is data scarcity, particularly for specific malware families and behaviors, which leads to imbalanced datasets and hinders classifier generalization. While synthetic data has been proposed as a remedy, the specific efficacy of Large Language Models (LLMs) in generating structured, effective malware records for training detection models remains underexplored. Existing literature lacks empirical evaluation on how LLM-generated synthetic Android malware influences detection performance when used for augmentation or as a standalone training source.

Methodology
This study utilizes the KronoDroid dataset, selected for its hybrid static–dynamic feature space, balanced malicious/benign splits, and human-interpretable features. The research focuses on three distinct malware families: BankBot (banking trojan), Locker/SLocker (ransomware), and Airpush/StopSMS (adware/SMS trojan).

The core methodology involves a three-stage process:

Data Preparation: The authors selected real device subsets from KronoDroid, maintaining a 1:1 ratio between malware and benign samples. Features were preprocessed to remove identifiers and metadata, imputing 'None' values to zero and filtering sparse columns, resulting in a final set of 387 numeric features.
LLM Fine-Tuning and Generation: The authors fine-tuned GPT-4.1-mini to generate structured malware records. To navigate safety filters regarding terms like "malware," feature names were sanitized (e.g., "malware" $\to$ $\to$ "app"). The team initially attempted CSV generation but encountered structural inconsistencies; consequently, they shifted to JSON/JSONL formats, which proved more stable for LLMs.
- Fine-tuning Strategy: Two configurations were tested. For BankBot and Locker/SLocker, 50 representative samples per family were used for one epoch. For Airpush/StopSMS, the set was increased to 150 samples over 3 epochs to test sensitivity to data volume and training depth.
- Prompt Engineering: Prompts were optimized to enforce structural realism, uniqueness, and correct labeling. Post-processing involved removing duplicates and addressing hallucinations (e.g., unrealistic system call patterns).
Experimental Evaluation: The study evaluated five classifiers (k-Nearest Neighbors, Decision Trees, Logistic Regression, Multilayer Perceptron, and Random Forest) across three scenarios:
- Real-only: Training on real malware and benign data.
- Real + Synthetic: Augmenting real malware with LLM-generated samples.
- Synthetic-only: Training exclusively on synthetic malware and real benign data, then testing on real malware.

Key Results
The experimental results, validated across 45 combinations (3 families $\times$ 5 classifiers $\times$ 3 scenarios), reveal distinct performance tiers:

Real-Only Training: Achieved near-perfect detection across all families. For BankBot, models like Logistic Regression and Random Forest achieved 1.000 accuracy and ROC AUC. Locker and Airpush families also showed high performance (Accuracy $\approx$ 0.97–0.99), confirming that sufficient real data yields excellent discrimination.
Real + Synthetic Augmentation: Adding LLM-generated samples preserved high performance but resulted in minor, consistent degradations compared to real-only training. For example, BankBot accuracy dropped slightly from 1.000 to ~0.99. Crucially, the False Positive Rate remained very low (<0.04), indicating that synthetic augmentation does not compromise specificity.
Synthetic-Only Training: Performance was mixed and highly dependent on the malware family and fine-tuning strategy:
- BankBot & Locker/SLocker: Models trained solely on synthetic data performed poorly when tested on real data. Accuracy ranged from 0.50 to 0.74, with extremely low recall (e.g., 0.05 for Locker), suggesting the synthetic data failed to capture the necessary distributional characteristics for these families.
- Airpush/StopSMS: This family showed significantly better generalization (Accuracy $\approx$ 0.80–0.89, Recall $\approx$ 0.61–0.79). This improvement correlated with the larger fine-tuning set (150 samples) and deeper training (3 epochs), suggesting that synthetic fidelity improves with increased data volume and training depth.

Key Contributions
The paper makes two primary contributions:

Augmentation Utility: It demonstrates that augmenting scarce real datasets with LLM-generated malware can enhance dataset size without significantly compromising detection accuracy or specificity.
Synthetic-Only Feasibility Limits: It provides empirical evidence that while LLMs can generate structurally consistent records, they are currently insufficient as a standalone training source for malware detection. The effectiveness of synthetic-only training is highly sensitive to the malware family characteristics and the depth of the fine-tuning process.

Significance and Claims
The authors position this work as a practical exploration of LLM utility in cybersecurity analytics rather than a breakthrough in raw detection metrics. The study claims that:

Synthetic data is a supplement, not a substitute: LLM-generated malware can effectively address data scarcity for rare families but cannot yet replace real-world data for training robust classifiers.
Methodological rigor is required: The success of synthetic generation depends heavily on prompt engineering, post-processing, and the specific fine-tuning regime (sample size and epochs).
Future potential: The findings suggest that increasing fine-tuning corpora and epochs for larger families could narrow the generalization gap. The authors propose that synthetic data is best applied as a complement for tasks like adversarial training, red-teaming, and benchmarking in environments where access to sensitive real-world datasets is limited.

In conclusion, the paper asserts that while LLMs offer a promising avenue for enriching malware datasets, current outputs lack the realism and diversity required for standalone detection models, necessitating a hybrid approach that prioritizes real data validation.

The Setup: The "Photo Album" and the "AI Artist"

The Three Experiments

The Big Takeaway

The Analogy of the "Cooking Class"

More like this