Lightweight And Privacy-Preserving Intrusion Detection For The Internet Of Things: A Systematic Review Of Datasets, Models, And Emerging Directions

Internet of Things devices now span consumer, industrial, and critical-infrastructure settings, and their growing numbers have expanded the attack surface available to adversaries, including botnet recruitment, denial-of-service, reconnaissance, and protocol-level exploitation. Prior surveys have established that IoT security requires solutions distinct from conventional enterprise network defense, given the constrained processing power, memory, and energy supply typical of IoT endpoints, and given the heterogeneity of IoT device populations and communication protocols. Intrusion detection is one of the central defensive mechanisms proposed for this setting, and a large body of work has focused specifically on making IDS models lightweight enough for on-device or near-edge deployment.

This review focuses on that lightweight IoT-IDS literature specifically, rather than IoT security broadly. Our motivation is that the field has grown quickly enough, across enough disconnected sub-threads, that a consolidated view of what has been tried, what results have been reported, and what remains open is useful both to newcomers and to researchers seeking to identify genuinely novel contributions rather than incremental replications. We restrict scope to detection (rather than prevention or response), to IoT/IIoT network and device traffic (rather than general-purpose network intrusion detection), and to approaches that explicitly report a resource-efficiency dimension (model size, inference latency, energy, or communication cost) alongside detection accuracy.

The remainder of this paper is organized as follows. Section 2 describes the review methodology, including search strategy and inclusion criteria. Section 3 surveys benchmark datasets. Section 4 reviews model-level lightweighting techniques. Section 5 covers federated and privacy-preserving approaches. Section 6 discusses explainability. Section 7 presents a comparative synthesis across the reviewed studies. Section 8 discusses open challenges and future directions, and Section 9 concludes.

2. REVIEW METHODOLOGY

This review follows a narrative-systematic approach informed by standard systematic-review search practice. Search terms combined IoT/IIoT with intrusion detection, lightweight, TinyML, edge, federated, and quantization/pruning, applied to major indexing sources (IEEE Xplore, ScienceDirect, SpringerLink, arXiv, and general web search) for work published primarily between 2019 and 2026, with emphasis on 2023–2026 publications to capture the current state of the field. Studies were included if they (a) targeted IoT or IIoT network/device traffic specifically, (b) proposed or evaluated a detection approach with an explicit resource-efficiency consideration (model size, latency, energy, communication cost, or an equivalent constrained-deployment framing), and (c) reported quantitative detection performance on at least one named benchmark dataset. Purely theoretical or position papers without an evaluated model, and papers addressing physical-layer or protocol-level security mechanisms unrelated to learned detection models, were excluded. Given the volume and rapid turnover of this literature, this review does not claim exhaustive coverage; it prioritizes recent (2023–2026), well-cited, and thematically representative work sufficient to characterize the current landscape and its gaps.

3. BENCHMARK DATASETS

Evaluation in this literature relies on a small number of recurring benchmark datasets. The Bot-IoT dataset provided an early realistic botnet-focused benchmark combining legitimate and simulated botnet traffic for network forensic analytics. The ToN_IoT dataset family, developed at UNSW Canberra Cyber, extended this by integrating IoT sensor telemetry, Linux and Windows OS audit logs, and network flow records from a multi-layer edge–fog–cloud testbed, covering nine attack categories with pronounced class imbalance (for example, man-in-the-middle traffic representing roughly 1,000 of over 200,000 total records in commonly used subsets). CICIoT2023, released by the Canadian Institute for Cybersecurity, was generated from a topology of 105 real IoT devices and captures 33 attack types across seven categories (DDoS, DoS, reconnaissance, web-based, brute force, spoofing, and Mirai-related botnet activity), and its authors explicitly note dataset-transferability analysis as an open direction. Additional datasets recurring in the reviewed literature include Edge-IIoTset and RT-IoT2022, both used to evaluate industrial and real-time IoT scenarios, and N-BaIoT, a botnet-focused dataset collected from commercial IoT devices infected with Mirai and BASHLITE. A recurring observation across dataset-overview studies is the lack of standardization in feature sets and attack taxonomies across these benchmarks, which complicates direct cross-study comparison and motivates the harmonization concerns discussed in Section 8.

4. MODEL-LEVEL LIGHTWEIGHTING TECHNIQUES

4.1 Feature Reduction and Classical Classifiers

A recurring strategy pairs dimensionality reduction with classical classifiers. Reported work combining principal component analysis with metaheuristic feature selection (particle swarm optimization) ahead of a support vector machine classifier reduced a 41-feature representation to 18 features on NSL-KDD while reporting 98.5% accuracy, illustrating that substantial feature-set compression can be achieved with limited accuracy cost. Comparative studies of decision trees, gradient-boosted ensembles (e.g., LightGBM), and compact neural networks on ToN_IoT have found tree-based classifiers competitive on raw F1-score (reported as high as 0.987 for a decision tree) but substantially larger in serialized size (hundreds of kilobytes) than compressed neural alternatives, illustrating that accuracy alone is an incomplete basis for comparison in this setting.

4.2 Pruning, Quantization, and Compact Deep Architectures

A second thread applies model compression — pruning, post-training or dynamic quantization, and architecture-level compaction — directly to deep models. Reported hybrid DNN-BiLSTM architectures combined with dynamic quantization and incremental PCA have achieved model sizes near 25–31 kilobytes while retaining accuracy above 93–99% depending on the benchmark (CIC-IDS2017 and CICIoT2023 respectively), and hybrid CNN-BiLSTM architectures explicitly designed for on-board IoT deployment have reported 97.28% binary and 96.91% multiclass accuracy on UNSW-NB15. Separately, pipelines combining pruning and quantization with generative oversampling (CTGAN) to address class imbalance have reported comparable accuracy to unpruned baselines while reducing training time by up to 94% and memory consumption by up to 90%. TinyML-optimized feedforward networks evaluated on ToN_IoT have reported F1-scores near 0.976 at 31 kilobyte model size with inference throughput near 120,000 packets per second, presented as a favorable balance point relative to larger classical baselines.

4.3 Spiking Neural Networks and Energy-Centric Approaches

A smaller thread targets energy consumption directly rather than treating it as a secondary consequence of model size. Spiking neural network formulations evaluated against conventional CNN baselines on NSL-KDD have reported up to 4.8× lower energy consumption at a moderate accuracy cost (approximately 75% accuracy and 72% F1 versus higher-accuracy CNN baselines), framing an explicit accuracy-versus-energy trade-off rather than optimizing accuracy alone.

4.4 TinyML and On-Device Deployment

TinyML-specific work extends compression techniques with deployment validation on constrained hardware and multi-dataset evaluation. Recent work applying feature-importance-based selection, quantization, pruning, and sparsification across tree-based and neural models has reported maintained accuracy across four benchmark datasets (CICIoT2023, Edge-IIoTset, CIC-IDS2017, and RT-IoT2022) after optimization, and a 2026 framework combining TinyML with graph attention networks, federated training, and explainability reported up to 95% accuracy across DDoS, advanced-persistent-threat, and zero-day attack scenarios evaluated on Edge-IIoTset, CICIoT2023, and RT-IoT2022, alongside a 20% reduction in federated communication overhead in a ten-node simulation.

5. Federated and Privacy-Preserving Approaches

A distinct and rapidly growing thread applies federated learning (FL) to IoT intrusion detection, motivated by the observation that centralizing raw traffic data from distributed IoT deployments raises privacy and bandwidth concerns. Systematic reviews of FL-based IDS have catalogued a broad range of architectures, and IoT-specific surveys have specifically examined privacy-preserving FL for intrusion detection. Reported IoT-focused FL-IDS work spans lightweight neural architectures combined with differential privacy and homomorphic encryption for real-time detection, mist- and edge-assisted FL architectures for heterogeneous IoT networks, and dynamic feature-fusion FL approaches targeted at vehicular network security. A consistent theme across this sub-literature is a distinction between two threat layers: external network intrusions, which FL-based IDS aim to detect, and internal model-layer attacks such as poisoning, backdoor insertion, and gradient leakage, which target the federated training process itself and are treated as a separate, compounding security concern. Non-IID data distribution across clients and class imbalance for rare attack types are repeatedly identified as open challenges specific to the federated setting, with proposed mitigations including cross-client aggregation adjustments and synthetic data generation for underrepresented classes.

6. Explainability in IoT Intrusion Detection

As lightweight and federated IDS models have matured, a parallel thread has emphasized explainability, motivated by the operational need for analysts to trust and act on model outputs rather than treat them as opaque alerts. Reported approaches include post-hoc explanation methods (e.g., LIME and SHAP) applied to multi-layer perceptron IDS models, attention-based interpretability built into graph neural network architectures for edge intrusion detection, and logic-based explainable frameworks proposed for adjacent network domains such as 5G. Explainability is increasingly reported as a module integrated alongside lightweight and federated components rather than as a standalone contribution, reflecting a trend toward multi-objective IDS designs that jointly target accuracy, efficiency, privacy, and interpretability.

7. Comparative Synthesis

Table 1 summarizes representative reported results across the studies reviewed in Sections 4–6, organized by technique, evaluation dataset, headline detection result, and the resource-cost dimension emphasized by each study. Figures are as reported by the original authors under their own experimental conditions; because evaluation datasets, preprocessing, and hardware platforms differ across studies, these figures should be read as illustrative of technique-level trends rather than as a directly comparable leaderboard.

Table 1. Representative reported results across lightweight IoT-IDS techniques (figures as reported in cited original studies; not independently reproduced).