Taming the Security-Energy Paradox: A Green AI Approach to Optimized Android Malware Detection

Shrinidhi Sridhar; Vikas K. Malviya

REVIEW 5 major objections 6 minor 31 references

Taming the Security-Energy Paradox: A Green AI Approach to Optimized Android Malware Detection

T0 review · 5 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read INT8 quantization can reduce Android malware-detection energy to 0.0189 mJ per inference while preserving accuracy.

desk verdict The classification and size benchmarks are fine, but the energy analysis is a restatement of latency with an unmeasured, inconsistent power coefficient, so the headline claim doesn't hold. read the letter →

arxiv 2607.20003 v1 pith:XFBF6RVN submitted 2026-07-22 cs.CR cs.AIcs.LG

Shrinidhi Sridhar , Vikas K. Malviya This is my paper

classification cs.CRcs.AIcs.LG

keywords AndroidmalwaredetectionINT8quantizationpost-trainingmultilayerperceptronenergy-efficientAITensorFlowLiteTUANDROMDDREBIN

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that the usual trade-off between strong Android malware detection and battery drain can be largely avoided by pairing shallow multilayer-perceptron models with post-training INT8 quantization. Across the TUANDROMD and DREBIN datasets, it reports that quantized 3- and 4-layer networks match or nearly match the accuracy of full-precision models while cutting model size by about 3.5 times and lowering estimated energy to as little as 0.0189 mJ per inference. The proposed mechanism is that smaller, lower-bit models finish inference faster, so the CPU spends less time in a high-power state and energy falls roughly with latency. If the estimates hold, developers get a concrete recipe for real-time background malware scanning that barely affects battery life. The energy figures are estimated, not measured on physical hardware, which is the paper's own stated limitation.

What carries the argument

The load-bearing ingredient is post-training quantization - converting a trained FP32 model's weights to 8-bit integers without retraining - applied to multilayer perceptrons and packaged as TensorFlow Lite models for on-device execution. Depth (3, 4, or 5 layers) and precision (FP32 vs INT8) are the two independent variables. Energy is estimated with E = P_cpu x t_enc + E_io, where P_cpu is a fixed processor power value and t_enc is per-inference latency, so lowering latency directly lowers reported energy; throughput and model size are the supporting metrics.

What would settle it

Run the same 3-layer INT8 TensorFlow Lite model on the TUANDROMD feature set inside a physical Android smartphone with battery instrumentation, measure average energy per 1,000 inferences, and compare to the paper's estimated 0.0189 mJ. If measured energy deviates by more than the 22% spread the paper reports across configurations, the fixed-P_cpu assumption is not transferable to real hardware.

Watch

Extended reading notes

Core claim

The paper's central claim is that post-training quantization from FP32 to INT8, combined with limited network depth, resolves the accuracy-energy tension in Android malware detection. Using 3-, 4-, and 5-layer MLPs on two public datasets, it reports that an INT8 3-layer network on TUANDROMD achieves 99.24% accuracy at 0.0189 mJ per inference, and an INT8 4-layer network on DREBIN achieves 98.63% accuracy at 0.0231 mJ; these are the configurations the paper recommends. Quantization reduces model size from 405.42 KB to 115.46 KB in the TUANDROMD case, and the paper ties energy consumption to the product of a constant CPU power and inference time. The conclusion is that strong security performa

Load-bearing premise

The load-bearing premise is that a single fixed CPU power value, applied to an x86_64 Android emulator, accurately represents a real phone's power draw during ML inference; if that value is wrong, every energy figure in the comparison, including the headline 0.0189 mJ, would shift.

Editorial extensions

If this is right

INT8 quantization would give near-lossless malware detection (accuracy loss of at most a few hundredths of a point in the reported tables) while shrinking model size by roughly 3.5x, easing memory pressure on low-end phones.
Because the paper reports a linear energy-latency relationship, any optimization that shortens inference time - not just quantization - should proportionally reduce battery cost for background security scans.
Throughput of about 47,000 inferences per second on the TUANDROMD 3-layer QNN suggests real-time continuous scanning is feasible without noticeable user-visible lag.
Developers can treat the identified configurations (3-layer INT8 for behavioral feature sets, 4-layer INT8 for manifest-based feature sets) as starting points for energy-aware malware-detector deployment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same quantization-and-depth trade-off likely transfers to other tabular Android feature sets, so a shallow INT8 model could become a general default for on-device scanning; this is an extrapolation beyond the two datasets tested.
The absolute 0.0189 mJ figure is probably less durable than the relative ordering of configurations, since the ordering follows from latency alone while the absolute figure depends on the unmeasured P_cpu value.
A testable follow-up suggested by the paper's own future-work list: check whether INT8 quantization shifts the calibration of predictive confidence; even if accuracy holds, security decisions based on score thresholds could change.
Because the experiments ran on an x86_64 Android emulator rather than an ARM phone, the 3.5x size reduction is hardware-independent but the energy advantage itself needs confirmation on physical devices with battery instrumentation.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a straightforward benchmark of INT8 post-training quantization for small MLPs on two Android malware datasets. The classification and model-size numbers are believable and consistent with prior work, but the energy analysis is not. The headline claim of 0.0189 mJ per inference comes from a single TUANDROMD configuration and is contradicted by the paper's own DREBIN table, where every INT8 model uses more energy than its FP32 counterpart.

What the paper does well: the experimental design is simple and transparent. They vary three depths, two precisions, and two datasets, report standard classification metrics, and make the emulator limitation explicit in the conclusion. The ~3.5x size reduction from FP32 to INT8 is what you'd expect, and the accuracy numbers (99%+ on TUANDROMD, 98.5% on DREBIN) are plausible. If you need a quick reference for TFLite latency of small MLPs on an x86 AVD, the throughput and latency columns have some empirical value.

The problem is the energy column. Eq. (1) from Hu et al. requires a measured or reported P_cpu and E_io. Neither is given. The paper states the AVD can't do battery instrumentation, which is fine for an admission, but then they apply the equation anyway. Dividing their energy by latency reveals implied P_cpu values around 0.99 mW for FP32 and 0.89 mW for INT8 models—different per model, with no explanation. That means the energy numbers are not a faithful application of their own equation; they're either arbitrary or the result of arithmetic inconsistency. And the DREBIN rows show quantization increasing energy, which is a direct contradiction of the abstract's blanket claim. The 0.0189 mJ figure is a single TUANDROMD 3-layer result that gets repeated in the abstract and conclusion as if it were general.

Who this is for: developers looking for a rough sense of model size and inference latency for tiny ML malware detectors on an Android emulator. That's a narrow audience, and they'd be better served by measuring on real hardware. The energy conclusions should not be cited until redone with actual power instrumentation.

As submitted, I would not accept this for peer review. The central quantitative claim is unsupported, and the fix requires new experiments, not just a rewrite. The classification findings are fine but too thin to stand alone. If the authors measure power on real devices and report their parameters, a revised version could be worth a look.

Referee Report

5 major / 6 minor

Summary. The paper reports an empirical comparison of FP32 and post-training-quantized INT8 MLP classifiers for Android malware detection on the TUANDROMD and DREBIN datasets. Three model depths (3, 4, and 5 layers) are compared in terms of classification metrics (accuracy, precision, recall, F1, MCC) and resource/efficiency metrics (model size, throughput, latency, and 'energy' computed from Eq. (1)). The claimed finding is that shallow INT8 models preserve high accuracy while reducing model size by about 3.5x and energy to 0.0189 mJ per inference, identifying a 3-layer QNN for TUANDROMD and a 4-layer QNN for DREBIN as optimal.

Significance. The paper addresses a genuine and timely problem: quantifying the trade-off between detection accuracy and on-device energy consumption in Android malware scanning. The classification results in Table I are plausible and suggest that INT8 quantization has little effect on detection quality for these MLP models, which is a useful observation. However, the paper's central contribution—the energy analysis—is not established. The energy values are derived from Eq. (1) with unmeasured coefficients, are internally inconsistent with that equation, and are contradicted for DREBIN by the paper's own data. Because the energy claims are the claimed novelty, the paper as written does not deliver its main result. The comparison of latency and model size is more credible, but those are not the metrics highlighted in the abstract.

major comments (5)

[§IV-C, Eq. (1), Table II] The energy numbers are not measurements. Eq. (1) requires P_cpu and E_io, but neither is reported or justified for neural-network inference; Section IV-C concedes the AVD has no physical battery instrumentation. The energy values are therefore rescaled latencies, and the 'linear energy-latency relation' in Section V-B1 is a direct consequence of the definition, not an empirical finding. No error bars or confidence intervals accompany the 10-run averages. The abstract's 0.0189 mJ headline and the conclusions in Section VI rest entirely on this unvalidated computation.
[Table II vs Eq. (1)] With fixed P_cpu and E_io, E/t_enc should be approximately constant unless E_io dominates. The table is not consistent with any single (P_cpu, E_io): FP32 rows cluster near 0.99 mW (e.g., TUANDROMD FP32 3-layer: 0.0238/0.0240 = 0.992 mW), while QNN rows cluster near 0.89 mW (e.g., TUANDROMD QNN 3-layer: 0.0189/0.0210 = 0.900 mW). Thus the implicit power coefficient changes between model classes without explanation. The reported energy numbers cannot have been produced by applying Eq. (1) with one measured P_cpu.
[Table II, DREBIN rows, abstract, §V-B2] The blanket claim that INT8 quantization reduces energy is contradicted by the DREBIN rows: each QNN configuration uses more energy than its FP32 counterpart (3-layer 0.0248 vs 0.0219 mJ; 4-layer 0.0231 vs 0.0224 mJ; 5-layer 0.0244 vs 0.0238 mJ) and has lower throughput. The 0.0189 mJ value is one TUANDROMD 3-layer point. §V-B2's statement that 'computational efficiency increased when the model precision was changed from FP32 to INT8' is therefore not supported as a general conclusion.
[Abstract vs Table I] The abstract claims 'more than 99.2% detection accuracy,' but the DREBIN results are all below 99.2% (maximum 98.63% for the identified optimal 4-layer QNN), and even TUANDROMD FP32 4-layer is 99.19%. The headline accuracy claim is therefore not representative of the paper's results, including the proclaimed optimal configuration for DREBIN.
[§IV-A] The MLP architectures are under-specified. Layer widths, optimizer, learning rate, batch size, epochs, and training details are not reported; only the number of layers is given. The 'optimal configurations' in §V-B3 cannot be reproduced or independently verified without these details or release of code/models.

minor comments (6)

[§II-D, Table II] The variable t_enc is defined in Eq. (1) as encryption/decryption time in the cited disk-encryption work, but Table II uses it as inference latency. The mapping should be justified and clearly labeled.
[References] Several references are not archival scholarly sources: [16] (GeeksforGeeks), [21] (ICML call for papers), [30] (KDD call for papers), and [31] (ResearchGate). The related-work claims should be grounded in peer-reviewed literature.
[Table I] There are formatting errors in the 5-layer rows, e.g., '599.31' should read '5 99.31'.
[§V-B2] The assertion that quantization reduces RAM consumption is not supported by any measured RAM metric in Table II or elsewhere.
[Tables I–II] Metrics are reported as point averages without standard deviations or statistical significance tests. Given the 5-fold cross-validation and 10-run protocol, dispersion should be reported.
[§V-B3] 'Optimal balance' is not defined quantitatively; no Pareto front or scalarized trade-off criterion is given. For DREBIN, the 4-layer QNN is selected despite having higher energy than the FP32 3-layer (0.0231 vs 0.0219 mJ), so the trade-off rationale is unclear.

Circularity Check

1 steps flagged · score 6.0 of 10

The energy analysis is not independent: Eq. (1) defines E as P_cpu·t_enc and P_cpu is never measured, so the headline energy saving is latency rescaling.

self definitional [Eq. (1) in Section II-D; applied in Section IV-C; energy column of Table II in Section V-B]
"EnergyErequired for a secure operation can be calculated as per the following equation [19]: E=P cpu ×t enc +E io (1) ... Since A VD environment doesn’t have the ability for physical battery instrumentation, we used the approach given by Hu et al. [19]. ... As per the data shown in the table, energy consumption (E) showed a linear relation with latency (t enc)."

The energy values are not measured: Eq. (1) sets E = P_cpu·t_enc + E_io, but neither P_cpu nor E_io is reported and Section IV-C says the AVD cannot do battery instrumentation. Therefore Table II's Energy column is, by construction, the t_enc column multiplied by an unstated constant. The central claim that INT8 quantization achieves 'a decrease in energy consumption to 0.0189 mJ per inference' is exactly the TUANDROMD 3-layer QNN latency 0.0210 ms relabeled as energy; no independent energy evidence is added. Moreover, a fixed P_cpu would make E/t_enc constant, yet the table implies P from about 0.89 to 0.996 mW across rows (0.0189/0.0210 vs 0.0238/0.0240), so the table is not even a consistent application of Eq. (1).

full rationale

The paper's accuracy and model-size results are self-contained empirical measurements from the chosen datasets and TFLite files; no circularity there. The circularity is confined to the energy contribution. The paper borrows Eq. (1) from Hu et al. [19], but that borrowing is not itself circular: the problem is that the formula is then used with no measured P_cpu (and no E_io), turning every energy number into a constant multiple of t_enc. The paper even states that energy 'showed a linear relation with latency,' confirming that the energy column carries no information beyond latency. Because the abstract's headline energy-saving number (0.0189 mJ) is simply the 3-layer QNN latency result rescaled, the energy conclusion reduces by construction to the latency measurement. The limitation paragraph in Section VI concedes only that 'emulators do not fully capture real-world hardware behavior,' which understates the issue: no physical battery instrumentation means no power value was measured at all. There is no load-bearing self-citation; the issue is the unmeasured constant in an externally borrowed equation. Score 6 rather than higher because classification performance and model-size reduction remain independent, non-circular contributions; the circularity affects the paper's distinctive energy claim but not its entire derivation.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central energy comparison rests on an unmeasured power model and emulator timing; classification comparisons rest on public datasets and unreported hyperparameters.

free parameters (3)

P_cpu (CPU power coefficient) = not reported; implied ~0.9–1.0 W from Table II
All energy values in Table II are computed as E = P_cpu × t_enc + E_io, but P_cpu is never stated or measured; back-calculating from Table II yields inconsistent values across rows.
E_io (I/O energy) = not reported
Equation (1) includes E_io, but the paper never describes its value or whether it is zeroed.
MLP layer widths and training hyperparameters = not reported
Model definitions (neurons/layer, epochs, optimizer, batch size) are omitted; they determine both accuracy and latency.

assumptions (3)

domain assumption Emulator latency (x86_64 host) is a valid proxy for real-device energy behavior.
Section IV-B/C uses AVD timing to compute energy; the paper itself concedes in Section VI that emulators do not capture real hardware behavior.
domain assumption Energy model E = P_cpu × t_enc + E_io (Hu et al.) applies to MLP inference on Android with constant P_cpu.
Section IV-C1 invokes Eq. 1 without justification that CPU power is constant across FP32/INT8 workloads.
domain assumption Datasets are already clean; labels/features are reliable.
Section IV-A relies on the two public datasets without leakage or class-balance analysis; TUANDROMD has 4465 samples vs DREBIN 7249, features differ, so cross-dataset validity is assumed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Taming the Security-Energy Paradox: A Green AI Approach to Optimized Android Malware Detection." pith.science (2026). https://pith.science/paper/XFBF6RVN

@misc{pith2026260720003,
  author       = {Pith},
  title        = {Pith review of: Taming the Security-Energy Paradox: A Green AI Approach to Optimized Android Malware Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XFBF6RVN}},
  note         = {Machine review of arXiv:2607.20003}
}

read the original abstract

An increase in advanced Android malware requires the use of deep learning models, which can run on Android devices. But there is a trade-off between security and energy use, as strong detection models can drain the battery of devices fast. This work tests different Multi-Layer Perceptron (MLP) model configurations to balance malware detection performance and energy efficiency. In this work, we compared standard FP32 models with optimized INT8 quantized neural networks with different model depths using TUANDROMD and DREBIN datasets for both classification performance and energy consumption. The results show that INT8 quantization reduces model size by about 3.5 times with a decrease in energy consumption to 0.0189 mJ per inference, while maintaining more than 99.2\% detection accuracy. We found that shallow quantized architectures, such as 3-layer and 4-layer QNNs, reduce energy costs by improving throughput and shortening the time of CPU operating in a high-power state. This work shows that efficient malware protection can be achieved on resource-constrained smartphones and provides a foundation for Green AI in mobile security.

Figures

Figures reproduced from arXiv: 2607.20003 by the authors.

**Figure 1.** Methodology performance of the models was evaluated using accuracy, precision, recall, F1-score, and Matthews Correlation Coefficient (MCC). This optimized pipeline made sure that the developed models not only perform better in terms of accuracy but also consume less power. B. Android Emulator Setup and Mobile Benchmarking This phase was devoted to evaluating the models developed in the previous phase on an Androi… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 1 linked inside Pith

[1]

https://ai-benchmark.com/ download

AI-Benchmark — ai-benchmark.com. https://ai-benchmark.com/ download. [Accessed 30-04-2026]

2026
[2]

https://android-developers.googleblog.com/ 2017/05/here-comes-treble-modular-base-for.html

Here comes Treble: A modular base for Android — android- developers.googleblog.com. https://android-developers.googleblog.com/ 2017/05/here-comes-treble-modular-base-for.html. [Accessed 30-01- 2026]

2017
[3]

https://source.android.com/docs/core/ota/modular-system

Mainline — Android Open Source Project — source.android.com. https://source.android.com/docs/core/ota/modular-system. [Accessed 30-01-2026]

2026
[4]

https://gs.statcounter.com/ os-market-share/mobile/worldwide

Mobile Operating System Market Share Worldwide — Statcounter Global Stats — gs.statcounter.com. https://gs.statcounter.com/ os-market-share/mobile/worldwide. [Accessed 18-03-2026]

2026
[5]

https://developer.android.com/guide/platform

Platform architecture — Android Developers — devel- oper.android.com. https://developer.android.com/guide/platform. [Accessed 30-01-2026]

2026
[6]

https://source.android.com/docs/core/power/mgmt

Power management — Android Open Source Project — source.android.com. https://source.android.com/docs/core/power/mgmt. [Accessed 30-01-2026]

2026
[7]

https://archive.ics.uci.edu/dataset/855/tuandromd+ (tezpur+university+android+malware+dataset)

Tuandromd dataset. https://archive.ics.uci.edu/dataset/855/tuandromd+ (tezpur+university+android+malware+dataset). [Accessed 30-04-2026]

2026
[8]

Agarwal et al

M. Agarwal et al. AI-Powered Android Malware Detection using Machine Learning. InProceedings of the IEEE/CVF, 2025

2025

Show all 31 references

[9]

Hasan al Nidawi, Tieng Wei Koh, Kareem Dawood, and Ammar Khaleel. Energy consumption patterns of mobile applications in android platform: A systematic literature review.Journal of Theoretical and Applied Information Technology, 95:6776–6787, 12 2017

2017
[10]

H. Anwar. ARENA: A tool for measuring and analysing the energy efficiency of Android apps.arXiv preprint arXiv:2510.01754, 2025

2025
[11]

The Drebin Dataset

Daniel Arp. The Drebin Dataset. https://drebin.mlsec.org/. [Accessed 30-04-2026]

2026
[12]

Standardized Approach for Evaluating AI Model Infer- ence Efficiency.MDPI Sensors, 25(3), 2025

CarbonTracker. Standardized Approach for Evaluating AI Model Infer- ence Efficiency.MDPI Sensors, 25(3), 2025

2025
[13]

A Survey of Quantization Techniques in Embedded AI Toolchains

DTU. A Survey of Quantization Techniques in Embedded AI Toolchains. Technical report, Technical University of Denmark, 2025

2025
[14]

On security and energy efficiency in android smartphones

Joao Ferreira, Bernardo Santos, Wellington Oliveira, Nuno Antunes, Bruno Cabral, and Jo ˜ao Paulo Fernandes. On security and energy efficiency in android smartphones. In2023 IEEE/ACM 10th Inter- national Conference on Mobile Software Engineering and Systems (MOBILESoft), pages...

2023
[15]

Model Quantization for Edge Environments.arXiv preprint arXiv:2504.03360, 2025

Friha et al. Model Quantization for Edge Environments.arXiv preprint arXiv:2504.03360, 2025

2025 arXiv
[16]

Beginner Friendly Machine Learning Projects 2025, 2025

GeeksforGeeks. Beginner Friendly Machine Learning Projects 2025, 2025

2025
[17]

File-based encryption, 2025

Google. File-based encryption, 2025. Accessed: 2026-01-30

2025
[18]

G ¨otzfried and T

J. G ¨otzfried and T. M ¨uller. Analysing android’s full disk encryption feature.Journal of Wireless Mobile Networks, Ubiquitous Computing, and Dependable Applications, 5:84–100, 03 2014

2014
[19]

Lui, Wenjun Hu, Xiaobo Ma, Jianfeng Li, and Xiao Liang

Yang Hu, John C.S. Lui, Wenjun Hu, Xiaobo Ma, Jianfeng Li, and Xiao Liang. Taming energy cost of disk encryption software on data-intensive mobile devices.Future Generation Computer Systems, 107:681–691, 2020

2020
[20]

Societal Considerations of Representation Learning, 2026

ICLR. Societal Considerations of Representation Learning, 2026

2026
[21]

Ignatov et al

A. Ignatov et al. ICML 2026 Call For Papers. International Conference on Machine Learning, 2026

2026
[22]

Green AI Energy-Efficient Machine Learning Model Selector

IJIRSET. Green AI Energy-Efficient Machine Learning Model Selector. International Journal of Innovative Research in Science, Engineering and Technology, 2025

2025
[23]

Klieder et al

W. Klieder et al. Evaluating Static Analysis Alerts with LLMs. CMU SEI Blog, 2024

2024
[24]

The android platform security model (2023)

Ren ´e Mayrhofer, Jeff Vander Stoep, Chad Brubaker, Dianne Hackborn, Bram Bonn ´e, G ¨uliz Seray Tuncay, Roger Piqueras Jover, and Michael Specter. The android platform security model (2023). Technical report, Cornell University, 2023

2023
[25]

Mehrotra, R

D. Mehrotra, R. Srivastava, R. Nagpal, and D. Nagpal. Multiclass clas- sification of mobile applications as per energy consumption.Journal of King Saud University - Computer and Information Sciences, 33(6):719– 727, 2021

2021
[26]

T. T. Nguyen et al. Time-and-Energy consumption offloading for mobile devices.ResearchGate, 2024

2024
[27]

Potlapally, S

N.R. Potlapally, S. Ravi, A. Raghunathan, and N.K. Jha. A study of the energy consumption characteristics of cryptographic algorithms and security protocols.IEEE Transactions on Mobile Computing, 5(2):128– 143, 2006

2006
[28]

Comparative energy efficiency analysis of mobile apps on android and ios platforms

Khoirur Rozikin, Joseph Santoso, Migunani Migunani, and Noorul Hassan S. Comparative energy efficiency analysis of mobile apps on android and ios platforms. pages 1–7, 08 2025

2025
[29]

Kraken.me mobile: The energy footprint of mobile tracking

Immanuel Schweizer, Roman B ¨artl, Benedikt Schmidt, Fabian Kaup, and Max M ¨uhlh¨auser. Kraken.me mobile: The energy footprint of mobile tracking. In6th International Conference on Mobile Computing, Applications and Services, pages 82–89, 2014

2014
[30]

Seo et al

J. Seo et al. AI for Sciences Track Call for Papers. KDD 2026, 2026

2026
[31]

Wang et al

S. Wang et al. A Practical and Accurate Battery Emulator for Android Smartphones.ResearchGate, 2025. 6

2025

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

https://ai-benchmark.com/ download

AI-Benchmark — ai-benchmark.com. https://ai-benchmark.com/ download. [Accessed 30-04-2026]

2026

[2] [2]

https://android-developers.googleblog.com/ 2017/05/here-comes-treble-modular-base-for.html

Here comes Treble: A modular base for Android — android- developers.googleblog.com. https://android-developers.googleblog.com/ 2017/05/here-comes-treble-modular-base-for.html. [Accessed 30-01- 2026]

2017

[3] [3]

https://source.android.com/docs/core/ota/modular-system

Mainline — Android Open Source Project — source.android.com. https://source.android.com/docs/core/ota/modular-system. [Accessed 30-01-2026]

2026

[4] [4]

https://gs.statcounter.com/ os-market-share/mobile/worldwide

Mobile Operating System Market Share Worldwide — Statcounter Global Stats — gs.statcounter.com. https://gs.statcounter.com/ os-market-share/mobile/worldwide. [Accessed 18-03-2026]

2026

[5] [5]

https://developer.android.com/guide/platform

Platform architecture — Android Developers — devel- oper.android.com. https://developer.android.com/guide/platform. [Accessed 30-01-2026]

2026

[6] [6]

https://source.android.com/docs/core/power/mgmt

Power management — Android Open Source Project — source.android.com. https://source.android.com/docs/core/power/mgmt. [Accessed 30-01-2026]

2026

[7] [7]

https://archive.ics.uci.edu/dataset/855/tuandromd+ (tezpur+university+android+malware+dataset)

Tuandromd dataset. https://archive.ics.uci.edu/dataset/855/tuandromd+ (tezpur+university+android+malware+dataset). [Accessed 30-04-2026]

2026

[8] [8]

Agarwal et al

M. Agarwal et al. AI-Powered Android Malware Detection using Machine Learning. InProceedings of the IEEE/CVF, 2025

2025