A cross-architecture detection and classification framework named CAMRA achieved a macro-averaged F1 score of 0.752 plus or minus 0.012 and an overall accuracy of 98.4 percent on Internet of Things malware, according to a study published in the International Journal of Computational Intelligence Systems. When evaluated in a decentralized setting across ten random seeds, the federated adaptation model reached a macro-averaged F1 score of 0.741 plus or minus 0.008 with 98.6 percent accuracy. Both central and federated implementations consistently outperformed every evaluated single-modality model and standard baseline.
The expansion of connected systems has substantially broadened the attack surface for malicious software designed to exploit security vulnerabilities across disparate hardware architectures. The authors reported that conventional static detection methods struggle against these threats. Past frameworks often relied on a single data modality or a specific device architecture, limiting detection across heterogeneous environments.
To overcome these limitations, the research team designed CAMRA as a unified framework combining multimodal telemetry, deep reinforcement learning, and federated learning. In their experimental evaluations, the investigators utilized the CIC-YNU-IoT-MAL-26 benchmark dataset. The proposed system captures and integrates three distinct operational modalities: network-level packet transmissions gathered from PCAP records, hardware-level system activity gathered via SAR metrics, and system-level execution sequences gathered through STRACE system call monitoring. These separate information streams are joined together using a cross-modal attention mechanism to capture interdependencies across the different observational levels.
The cross-modal attention architecture operates across four distinct hardware platforms evaluated in the study: ARM, MIPS, MIPSEL, and x86 architectures. To enable practical deployment across decentralized devices, the framework incorporates federated learning algorithms. Under this scheme, a centrally trained detection model is adapted across architecture-specific client nodes. This federated adaptation allows local client devices representing different processor families to tune their classification models while keeping sensitive raw operational data strictly on local machines, preserving data locality across diverse nodes.
To tackle class imbalance across the malware samples, CAMRA incorporates a deep reinforcement learning controller that dynamically adjusts per-class loss weights throughout the training phase. Subsequent ablation studies and paired statistical tests conducted by the research team indicated that the frequency-based class weights and the designated loss function served as the primary drivers of model performance. The statistical tests further demonstrated that the adaptive reinforcement learning controller performed on par with fixed class weighting schedules rather than providing a significant statistical advantage over them.
Holdout validation experiments highlighted both the detection strengths and the operational constraints of the framework under unfamiliar deployment conditions. When evaluated on unseen processor architectures and an unseen malware family, CAMRA successfully recognized the samples as malicious. However, the holdout trials also identified clear limits to generalizability, as exposure to an unseen ARM architecture generated false alarms on benign software profiles. These results show that while multimodal fusion and federated adaptation enhance cross-architecture detection, unfamiliar hardware variations still trigger false alarms on benign data.
Subject of Research: Cross-architecture multimodal Internet of Things malware detection
Article Title: CAMRA: Cross-Architecture Multimodal Reinforcement Learning with Federated Adaptation for IoT Malware Detection and Classification
Article References: Gourkhede, A., Cajla, A., Chhabra, G. S., Verma, S., & Mahapatra, B. (2026). CAMRA: Cross-Architecture Multimodal Reinforcement Learning with Federated Adaptation for IoT Malware Detection and Classification. International Journal of Computational Intelligence Systems. https://doi.org/10.1007/s44196-026-01669-2
Image Credits: AI Generated
DOI: 10.1007/s44196-026-01669-2
Keywords: Internet of Things, IoT malware detection, malware classification, cross-architecture learning, multimodal data fusion, deep reinforcement learning, federated learning, cross-modal attention, CIC-YNU-IoT-MAL-26 dataset, network traffic PCAP, hardware metrics SAR, system call STRACE, ARM architecture, MIPS architecture, MIPSEL architecture, x86 architecture, class imbalance, macro-F1 score
News Source: Denise Maddox. (November 18, 2026). Multimodal Federated Framework Detects and Classifies IoT Malware Across Diverse Architectures. Scienmag.



