💻 computer science

TAH-GAN: Time-Aware Hybrid Generative Adversarial Network for Robust Android Malware Evasion

This paper introduces TAH-GAN, a time-aware hybrid generative adversarial network that integrates Fourier temporal encoding and dual-gate quality control to generate future-aligned Android malware samples, effectively addressing temporal concept drift and achieving superior evasion rates and temporal coherence compared to existing non-temporal baselines.

Original authors: Daniel Jeremiah, Jalil Md D, Husnain Rafiq

Published 2026-09-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Daniel Jeremiah, Jalil Md D, Husnain Rafiq

Original paper licensed under CC BY 4.0 (https://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

In the digital world of mobile phones, a constant, silent war is fought between the software that protects our devices and the malicious programs designed to break them. For years, security experts have relied on machine learning, a form of artificial intelligence that learns to spot threats by studying thousands of examples of bad software. However, this defense has a fundamental weakness: it is stuck in the past. These security systems are trained on data from previous years, but the threats they face are constantly changing. As new phone operating systems are released and hackers invent new tricks, the relationship between what a program looks like and whether it is dangerous shifts over time. This phenomenon, known as concept drift, means that a security guard trained on last year's criminals might fail to recognize a criminal who has changed their appearance and methods today. The challenge for researchers is not just to catch these new threats, but to understand how to build defenses that can anticipate them before they become widespread.

A team of researchers has developed a new tool designed to solve this specific problem of time and change in mobile security. They created a system called TAH-GAN, which stands for Time-Aware Hybrid Generative Adversarial Network. To understand what this system does, imagine a security team trying to prepare for a future attack. Instead of waiting for real hackers to strike, they need a way to generate fake, but realistic, examples of what those future attacks might look like. Previous attempts to create these fake examples often failed because they treated the data as static, ignoring the fact that the digital landscape changes every day. The new system, however, is built with a specific awareness of time. It does not just learn what malware looks like today; it learns how malware evolves. By feeding the system information about when data was collected, the researchers enabled it to generate samples that align with the future, creating a bridge between the past data security systems know and the future threats they must face.

The researchers tested this new approach using a massive collection of Android applications, separating the data strictly by time. They trained their system on older applications and then asked it to generate new, evasive examples that would appear in later years. The results were striking. When they compared their time-aware system to other existing methods that ignored the element of time, the difference was clear. The older methods produced fake samples that looked strange or inconsistent when viewed through the lens of time, much like a person wearing clothes from 1990 but carrying a smartphone from 2024. In contrast, the new system produced samples that were indistinguishable from real, future malware. In their tests, nearly all of the generated samples, specifically 99.8 percent, successfully passed two critical checks: they looked real enough to fool a detector, and they were successfully classified as safe by the security system they were trying to trick. This high success rate held true even when the researchers tested the generated samples against different types of security software that the system had never seen before, proving that the attack was not just a fluke against one specific program.

The study also revealed that the element of time was the most critical factor in this success. When the researchers removed the time-aware component from their system, the quality of the generated samples dropped significantly. The system without time awareness produced samples that were coherent in a general sense but failed to capture the specific patterns of a future era. This finding suggests that for security systems to remain effective, they cannot simply learn from a static database; they must understand the rhythm of change. The researchers used a mathematical technique involving waves to encode the passage of time, allowing the system to smooth out the transitions between years and predict how features would shift. This approach allowed them to create a dataset of over 1.7 million high-quality, time-aligned samples that can be used to train future security systems to be more robust.

Ultimately, this work provides a new way to stress-test mobile security. By generating realistic, future-aligned examples of malware, defenders can identify weaknesses in their systems before real attackers exploit them. The researchers made their code and the massive dataset of generated samples available to the public, allowing other scientists to verify the results and build upon them. While the system operates on the level of data features and does not yet create fully installable applications, it represents a significant step forward in understanding how to defend against threats that are constantly evolving. The study concludes that ignoring the dimension of time leaves security systems vulnerable, but by integrating time directly into the generation process, it is possible to create defenses that are ready for what comes next.

Technical Summary: TAH-GAN for Robust Android Malware Evasion

Problem Statement
Android malware detectors based on machine learning suffer from temporal concept drift, where models trained on historical data fail to detect novel variants emerging in future operational periods. Empirical evidence indicates that when training and test data are strictly separated by time, classifier accuracy can drop by up to 46 percentage points. While adversarial machine learning has been used to generate evasion samples, existing Generative Adversarial Networks (GANs) treat feature vectors as static, context-free representations. Consequently, they produce temporally incoherent samples (e.g., combining permission patterns from one year with API calls from another) that fail under temporally isolated evaluation. There is a critical gap in the literature: no prior framework generates future-aligned, temporally coherent adversarial malware samples suitable for drift-aware defensive retraining.

Methodology: TAH-GAN Framework
The authors propose TAH-GAN (Time-Aware Hybrid Generative Adversarial Network), a framework designed to generate adversarial samples that are realistic, evasive, and temporally coherent. The system operates on a temporally augmented representation $v(x, t) = [x \parallel \phi(t)]$ , where $x$ is the malware feature vector and $\phi(t)$ is a Fourier-encoded temporal condition.

Fourier Temporal Encoding: To address the static nature of existing generators, TAH-GAN conditions the generator $G$ on a 17-dimensional temporal vector $\phi(t)$ . This vector combines:
- Eight harmonics of sine and cosine functions derived from the normalized calendar year $\tau_y(t)$ , capturing periodic variations.
- A normalized Unix timestamp $\tau_u(t)$ , preserving monotonic progression.
  This encoding allows the generator to synthesize feature distributions aligned with specific target time periods.
Architecture and Training:
- Generator ( $G$ ): A 4-layer MLP that takes latent noise $z$ and the temporal condition $\phi(t)$ to produce a synthetic feature vector $\tilde{x}$ .
- Discriminator ( $D$ ): A Wasserstein GAN with Gradient Penalty (WGAN-GP) critic that evaluates the realism of the augmented vector $v(\tilde{x}, t)$ . It provides a realism objective to ensure the generated samples resemble real malware distributions.
- Classifier ( $C$ ): A pre-trained Multi-Layer Perceptron (MLP) used as a frozen target. It provides an evasion objective by calculating the probability that a sample is malicious. The generator is trained to minimize this probability (i.e., maximize evasion) while satisfying the discriminator's realism constraint.
- Loss Function: The generator minimizes a combined loss: $L_G = -\lambda_R E[D(v_g)] + \lambda_E E[\sigma(C(v_g))]$ , balancing realism and evasion.
Two-Stage Quality Control (QC):
Before releasing a full batch of generated samples, TAH-GAN employs a pilot batch test. A checkpoint is only released if the pilot batch meets two simultaneous thresholds:
- Realism Gate: At least 98% of samples satisfy $D(v) \geq 0$ .
- Evasion Gate: At least 98% of samples satisfy $\sigma(C(v)) < 0.5$ .
  This ensures that the final dataset consists of samples that are both realistic and evasive, rather than just averaging marginal success rates.

Key Contributions

First Time-Aware GAN for Malware: TAH-GAN is the first GAN-based framework to incorporate Fourier temporal encoding specifically for generating future-aligned adversarial malware under temporally isolated evaluation settings.
Drift-Aware Retraining Support: The framework produces samples that serve as "hard negatives" for retraining detectors, explicitly addressing the temporal drift problem rather than treating it as a downstream modeling issue.
Dual-Gate Quality Control: The introduction of a two-stage QC workflow (realism + evasion) ensures high-fidelity generation, rejecting samples that fail either constraint.
Open Resources: The authors release the implementation code, experiment scripts, and a dataset of 1,757,645 frozen, quality-controlled outputs.

Experimental Results
The framework was evaluated on the KronoDroid (dynamic system calls) and AndroMD (static features: MNF, KeyCount, ZeroOne) datasets using temporally isolated train-test splits.

Performance vs. Baselines: TAH-GAN significantly outperformed baselines (CTGAN, DoPGAN, and a standard EGAN).
- Balanced Gate Score (BGS): TAH-GAN achieved 99.7%, compared to 88.5% for CTGAN and 87.2% for DoPGAN.
- Joint Gate-Pass Rate (JGR): A stricter recomputation on all frozen outputs showed 99.816% of samples passed both realism and evasion gates simultaneously.
- Temporal Coherence: TAH-GAN achieved a temporal coherence score of 0.91, whereas non-temporal baselines (CTGAN, DoPGAN) scored only 0.12.
Evasion Effectiveness: When injected into a baseline classifier, TAH-GAN reduced the detection rate (recall) from ~98.1% to 0.5% on average.
Transferability: Samples generated to evade an MLP target achieved 82–87% evasion rates against unseen architectures (CNN1D, LSTM, ResNet), demonstrating architecture-agnostic effectiveness.
Ablation Studies: Removing the temporal encoding reduced the BGS by 2.0 percentage points, confirming the necessity of Fourier conditioning. Removing quality control caused the largest drop (5.0 percentage points).

Significance and Claims
The paper claims that TAH-GAN exposes temporally evolving decision-boundary weaknesses in malware detectors. By generating samples that are aligned with the distribution of future malware, the framework provides a robust mechanism for stress-testing and retraining defenses against concept drift.

The authors position TAH-GAN not as a tool to defeat production antivirus stacks (which often use signature-based or hybrid approaches), but as a security evaluation tool for ML-based detectors. The generated dataset serves as a reproducible resource for researchers to study drift-aware retraining and improve the robustness of future detection systems. The work highlights that without temporal conditioning, adversarial generators produce incoherent samples that fail to represent the true threat landscape of evolving malware.

Technical Summary: TAH-GAN for Robust Android Malware Evasion

More like this