Federated drift-resilient android malware detection through lightweight bytecode imaging

Android malware evolves continuously, causing machine-learning-based detectors trained on historical data to lose effectiveness against newer threats. At the same time, collaborative malware detection may be constrained by organisational and operational requirements that limit centralized pooling of raw application data. Temporal robustness and federated malware detection have largely been studied separately, leaving their combined behaviour under resource-constrained deployment insufficiently understood. This study presents a federated Android malware detection framework based on an Era-as-Client formulation, in which temporally distinct application cohorts are treated as separate federated clients. A lightweight convolutional neural network operates on $$32\times32$$ grayscale representations of Android DEX bytecode. Experiments use 20,000 real-world AndroZoo applications divided into Era-OLD (2013–2017) and Era-NEW (2019–2023), with evaluation performed over five random seeds. Direct OLD-to-NEW transfer reduces accuracy from 96.5% to 91.9%, with malware recall decreasing to 85.1%. This degradation is accompanied by measurable changes in the bytecode-image distribution: mean normalized pixel intensity increases from 0.292 to 0.326, image entropy from 6.18 to 6.44 bits, and active-pixel ratio from 0.735 to 0.794. A Kolmogorov–Smirnov test further confirms a significant temporal distribution difference ( $$D=0.20$$ , $$p < 10^{-10}$$ ). To separate the effect of contemporary labelled-data access from federated optimization, the Era-as-Client model is compared with matched Centralized Pooled and Sequential Adaptation controls. The federated model achieves 96.2% accuracy, 97.8% precision, 94.5% recall, 96.1% F1-score, 0.990 AUC, and 0.989 average precision. Under identical OLD+NEW training-data access, Centralized Pooled training reaches 96.6% accuracy and Sequential Adaptation reaches 96.1%. Thus, the federated model remains within 0.4 percentage points of centralized pooling, and its principal benefit is raw-data locality and decentralized collaborative training rather than superior predictive accuracy. A strict forward-in-time evaluation further reserves all 2022–2023 applications for final testing while restricting training, validation, and adaptation to data available through 2021. Under this setting, the OLD-only model achieves 90.4% accuracy, whereas the Federated Era-as-Client, Centralized Pooled, and Sequential Adaptation settings achieve 94.4%, 94.9%, and 94.2%, respectively, confirming the benefit of more recent training information on a genuinely later unseen period. Multi-client experiments additionally evaluate client-count scaling, non-IID partitioning, and partial participation. Accuracy remains 95.7% with ten clients under fixed local-data volume, 94.2% under strong Dirichlet heterogeneity ( $$\alpha=0.1$$ ), and 95.1% with 30% client participation. The model contains 48,662 trainable parameters, occupies 0.186 MB in FP32 form, and achieves an average CPU inference latency of 5.6 ms. Although $$128\times128$$ inputs improve accuracy by 1.0 percentage point, the selected $$32\times32$$ representation provides 2.6 $$\times$$ lower inference latency, supporting its use for latency-constrained deployment. Overall, the study provides quantitative evidence of temporal distribution shift in bytecode imagery, matched controls for interpreting federated adaptation, strict forward-in-time validation, and multi-client evaluation under controlled federated heterogeneity. The federated design provides raw-data locality by avoiding centralized APK pooling; standard FedAvg is not treated as providing a formal privacy guarantee without additional mechanisms such as differential privacy or secure aggregation.

Authors

Institutions

Publication Details

Journal
Discover Computing
Published
2026-11-17
DOI
https://doi.org/10.1007/s10791-026-10818-2
Primary Topic
Advanced Malware Detection Techniques
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Federated drift-resilient android malware detection through lightweight bytecode imaging

Rijvan Beg, Deepak Singh Tomar, Amit Patel, R. K. Pateriya
Discover Computing
Advanced Malware Detection Techniques
article

Federated drift-resilient android malware detection through lightweight bytecode imaging

Rijvan Beg, Deepak Singh Tomar, Amit Patel, R. K. Pateriya
article en

Abstract

Android malware evolves continuously, causing machine-learning-based detectors trained on historical data to lose effectiveness against newer threats. At the same time, collaborative malware detection may be constrained by organisational and operational requirements that limit centralized pooling of raw application data. Temporal robustness and federated malware detection have largely been studied separately, leaving their combined behaviour under resource-constrained deployment insufficiently understood. This study presents a federated Android malware detection framework based on an Era-as-Client formulation, in which temporally distinct application cohorts are treated as separate federated clients. A lightweight convolutional neural network operates on $$32\times32$$ grayscale representations of Android DEX bytecode. Experiments use 20,000 real-world AndroZoo applications divided into Era-OLD (2013–2017) and Era-NEW (2019–2023), with evaluation performed over five random seeds. Direct OLD-to-NEW transfer reduces accuracy from 96.5% to 91.9%, with malware recall decreasing to 85.1%. This degradation is accompanied by measurable changes in the bytecode-image distribution: mean normalized pixel intensity increases from 0.292 to 0.326, image entropy from 6.18 to 6.44 bits, and active-pixel ratio from 0.735 to 0.794. A Kolmogorov–Smirnov test further confirms a significant temporal distribution difference ( $$D=0.20$$ , $$p < 10^{-10}$$ ). To separate the effect of contemporary labelled-data access from federated optimization, the Era-as-Client model is compared with matched Centralized Pooled and Sequential Adaptation controls. The federated model achieves 96.2% accuracy, 97.8% precision, 94.5% recall, 96.1% F1-score, 0.990 AUC, and 0.989 average precision. Under identical OLD+NEW training-data access, Centralized Pooled training reaches 96.6% accuracy and Sequential Adaptation reaches 96.1%. Thus, the federated model remains within 0.4 percentage points of centralized pooling, and its principal benefit is raw-data locality and decentralized collaborative training rather than superior predictive accuracy. A strict forward-in-time evaluation further reserves all 2022–2023 applications for final testing while restricting training, validation, and adaptation to data available through 2021. Under this setting, the OLD-only model achieves 90.4% accuracy, whereas the Federated Era-as-Client, Centralized Pooled, and Sequential Adaptation settings achieve 94.4%, 94.9%, and 94.2%, respectively, confirming the benefit of more recent training information on a genuinely later unseen period. Multi-client experiments additionally evaluate client-count scaling, non-IID partitioning, and partial participation. Accuracy remains 95.7% with ten clients under fixed local-data volume, 94.2% under strong Dirichlet heterogeneity ( $$\alpha=0.1$$ ), and 95.1% with 30% client participation. The model contains 48,662 trainable parameters, occupies 0.186 MB in FP32 form, and achieves an average CPU inference latency of 5.6 ms. Although $$128\times128$$ inputs improve accuracy by 1.0 percentage point, the selected $$32\times32$$ representation provides 2.6 $$\times$$ lower inference latency, supporting its use for latency-constrained deployment. Overall, the study provides quantitative evidence of temporal distribution shift in bytecode imagery, matched controls for interpreting federated adaptation, strict forward-in-time validation, and multi-client evaluation under controlled federated heterogeneity. The federated design provides raw-data locality by avoiding centralized APK pooling; standard FedAvg is not treated as providing a formal privacy guarantee without additional mechanisms such as differential privacy or secure aggregation.

Discover ComputingVol. 29(1)
Maulana Azad National Institute of Technology (IN)
Maulana Azad National Institute of Technology
Advanced Malware Detection Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon