Privacy-Aware and Resource-Efficient Split Learning for IoT Botnet Detection: A Multi-Dataset Experimental and Systems Evaluation
TL;DR
A traffic-constrained multi-client split-learning intrusion-detection system evaluated on BoT-IoT, N-BaIoT, and CIC-IDS2017, which identifies a practical accuracy–communication–energy–privacy operating point for constrained IoT clients.
Abstract
Internet of Things (IoT) botnet detection requires high predictive performance, low client-side resource demands, and limited exposure of raw traffic data. This paper presents a traffic-constrained multi-client split-learning (TC-SL) intrusion-detection system evaluated on BoT-IoT, N-BaIoT, and CIC-IDS2017 using eight clients and one edge server. TC-SL is positioned as a resource-constrained engineering adaptation of standard split learning: it profiles candidate cut layers, selects the highest-performing feasible cut under an explicit communication budget, and trains by exchanging cut-layer activations and gradients while raw records remain local. Centralized, federated, split-learning, and SplitFed-v1 models were compared with matched partitions and optimization budgets. Split learning achieved macro F1 scores of 98.88%, 98.42%, and 97.51% on the three datasets, respectively, with a mean macro F1 of 98.27%, compared with 98.52% for centralized learning, 97.84% for federated learning, and 98.39% for SplitFed. Its pooled ROC-AUC and average precision were 0.985 and 0.980. Within the fixed 10-epoch systems workload, the selected L4 cut logged 95 MB of bidirectional communication per epoch and 0.95 GB in total, versus 405 MB per epoch and 4.05 GB for the configured federated-learning baseline; these byte totals are fixed-workload measurements rather than communication-to-target-accuracy claims. The early split minimized cumulative client energy at 248 J, whereas the middle split minimized total-system energy at 780 J. Across L1–L6, reconstruction NRMSE increased from 0.18 to 0.71, membership-inference AUC decreased from 0.71 to 0.53, label- and attribute-inference success decreased from 0.84 and 0.76 to 0.55 and 0.54, inversion success decreased from 75% to 20%, and poisoning-induced macro-F1 degradation decreased from 6.8 to 3.9 percentage points. The results identify a practical accuracy–communication–energy–privacy operating point for constrained IoT clients.
