Intrusion Detection for Agricultural IoT Networks: An Empirical Comparison of Supervised and Unsupervised Machine Learning Approaches

IoT in Smart Agriculture

Smart agriculture leverages a complex, multi-layered architecture—comprising perception, network, edge, fog, and cloud layers—to manage agricultural operations efficiently [1], [4]. This paradigm relies on the integration of smart sensing technologies, including soil moisture sensors, weather stations, automated irrigation, and livestock tracking systems. In precision irrigation, soil moisture sensors transmit real-time data to automated controllers that manage water distribution. Livestock monitoring systems utilize wearable sensors to track animal health and behavior. Greenhouse automation integrates environmental sensors with actuators to regulate temperature, humidity, and light conditions.

This architecture relies on diverse communication protocols, including long-range, low-power solutions like LoRaWAN for wide-area field monitoring, ZigBee for mesh networks within greenhouses, and lightweight application-level protocols like MQTT and CoAP. This protocol heterogeneity creates significant security complexity, as each layer and protocol presents different attack vectors. Edge computing has become essential, providing local data processing that alleviates load on centralized cloud servers and enables instantaneous response to sensor data.

Cybersecurity Threats in Agricultural IoT

AgriIoT networks are increasingly exposed to a broad spectrum of cyber threats, exacerbated by the lack of standardized security protocols and the presence of heterogeneous, often unpatched, devices [3], [5]. Common threats include DoS/DDoS attacks, malware and ransomware, botnet participation, spoofing, man-in-the-middle (MitM) attacks, and side-channel vulnerabilities [4]. A false data injection attack targeting soil moisture sensors could manipulate readings to trigger unnecessary irrigation—leading to crop waterlogging—or suppress irrigation entirely, destroying valuable harvests. The IT/OT convergence allows attackers to bridge office-level breaches to industrial-level operational disruption, a physical-cyber intersection unique to the agricultural context.

Machine Learning for Intrusion Detection

Supervised models, such as Random Forest and Support Vector Machines, excel in classifying known attack types using labeled datasets [8], [10]. Unsupervised models, such as Autoencoders, detect anomalies by modeling the baseline behavior of the network, making them suitable for identifying previously unseen or zero-day attacks [11], [12]. Unsupervised Autoencoders are trained to reconstruct input data; their effectiveness is measured by the reconstruction error—the difference between the original input and the reconstruction. A high reconstruction error acts as an anomaly threshold. In security systems there is a fundamental trade-off between the false-positive rate and recall: a system configured for high recall captures more attacks but at the cost of more false positives.

Security Standards and Frameworks for IoT

The proliferation of IoT devices has prompted the development of several cybersecurity frameworks, most notably the NIST IoT Cybersecurity Capabilities Baseline [13], the ETSI EN 303 645 standard for consumer IoT [14], and the IEC 62443 series for industrial automation and control systems [15]. These frameworks provide guidelines for secure device configurations, robust authentication, data encryption, and vulnerability management. However, implementing these standards in AgriIoT presents profound challenges. These guidelines were designed with enterprise IT environments in mind, assuming computational resources unavailable in low-power agricultural sensor networks. Low-power sensors, often operating on battery power for months in remote locations, cannot sustain the communication overhead required for heavy security protocols. Standardized update procedures fundamental to IEC 62443 compliance are often impossible due to intermittent connectivity. Consequently, AgriIoT deployments often prioritize functional availability over strict cybersecurity protocols, inadvertently leaving critical agricultural infrastructure vulnerable.

Intrusion Detection Systems for IoT

The development of efficient NIDS for IoT is a primary security objective, with numerous studies focusing on dimensionality reduction, feature selection, and lightweight model design [8], [16]. Bella et al. [17] proposed an efficient IDS utilizing a CNN-decision forest combination, achieving high detection accuracy and rapid inference. Li et al. [16] compared feature selection versus feature extraction methods, providing actionable insights for optimizing NIDS performance. While effective for general IoT, such models often rely on centralized processing and lack the context-awareness required to distinguish benign agricultural sensor fluctuations from malicious anomalies.

Cybersecurity in Smart Agriculture

Research dedicated to smart agriculture has emphasized holistic cybersecurity frameworks. Ali et al. [4] provided a systematic synthesis of AI applications in AgriIoT cybersecurity, documenting nearly 30 specific threat vectors. Thilakarathne et al. [18] proposed a deception-based threat-intelligence platform using honeypots to monitor real-world attack patterns. Pasca et al. [7] developed a vulnerable-by-design framework for generating labeled datasets integrating both cyber threats and sensor faults. A synthesis of these studies reveals that while the threat landscape for smart agriculture is maturing rapidly, defensive measures remain critically under-resourced, with few academic efforts offering robust, field-validated solutions tailored to the unique operational demands of modern farms.

Machine Learning Methods for Anomaly Detection

Maseer et al. [10] conducted a systematic benchmarking of ten algorithms on CICIDS2017, finding that ensemble methods like Random Forest provided the most stable and accurate results. Meidan et al. [11] presented N-BaIoT, employing deep autoencoders to detect botnet traffic in IoT networks by modeling normal behavior for individual device types. While Maseer et al. focus on benchmarking diverse algorithms for multi-class classification, Meidan et al. provide a specialized unsupervised approach for zero-day botnet detection through learned normal traffic reconstruction.

Benchmark Datasets

Benchmark datasets are essential for evaluating IDS performance. While NSL-KDD has historically served as a foundation, its efficacy in evaluating modern IoT traffic patterns is severely limited [8], [9]. Panigrahi [19] provided an extensive analysis of CICIDS2017, highlighting the need for rigorous preprocessing to handle class imbalance, scattered presence across multiple files, and missing values. The dataset covers diverse attack types including DoS, DDoS, PortScan, Infiltration, Web Attacks, and Botnet infiltration, making it the most appropriate contemporary benchmark for this study.

Dataset

This study utilizes the CICIDS2017 dataset, a comprehensive, realistic benchmark containing diverse network traffic including modern DDoS, DoS, infiltration, and web attacks. It comprises 2,827,876 total records with 78 network-traffic features, providing a rich, high-dimensional space for machine-learning evaluation. The flows are labeled by attack type, enabling both binary (normal vs. malicious) and multi-class classification.

Research Gap

Despite advancements, several critical gaps remain. First, there is a reliance on outdated datasets that fail to represent modern heterogeneous AgriIoT traffic [8], [9]. Second, existing IDS methodologies are tested in idealized laboratory settings, leaving a vacuum in field-validated results on actual AgriIoT hardware [6], [7]. Third, there is insufficient research into interactions between cyber-attacks and physical sensor failures. Fourth, there is a need for resource-efficient IDS models balancing detection performance with the constraints of agricultural sensor networks in developing contexts [5], [8]. This study addresses these gaps by evaluating both supervised and unsupervised approaches on CICIDS2017, exploring the trade-offs between accuracy, computational demands, and label availability.

Preprocessing

The pipeline involved cleaning the raw parquet files, handling missing values, and splitting data to prevent leakage. The supervised training set combined normal samples with attack samples, yielding approximately 100,000 samples for computational efficiency while maintaining representative attack patterns. Min–Max scaling mapped all numeric features into the [0, 1] range to prevent features with larger dynamic ranges from dominating gradient updates. Splitting was stratified to ensure proportional class representation.

Models

Five machine-learning models were evaluated:

1. Random Forest — an ensemble of decision trees that reduces overfitting via feature randomness and bootstrap sampling, capturing complex non-linear relationships in network traffic.
2. Decision Tree — a non-parametric method that recursively partitions the feature space by information gain, offering high interpretability for security analysts.
3. Logistic Regression — a linear baseline that illustrates the limitations of linear models on the non-linear patterns of modern intrusions.
4. Autoencoder — an unsupervised network (78 → 128 → 64 → 16 → 64 → 128 → 78) trained to reconstruct its input; the reconstruction error serves as the anomaly metric.
5. Variational Autoencoder (VAE) — a generative model learning the probability distribution of the input within a latent space governed by a prior, providing a probabilistic framework for anomaly detection.

Evaluation Metrics

Models were evaluated using five performance indicators:

• Precision — the ratio of true positive predictions to total positive predictions, reflecting the model's ability to minimize false alarms.
• Recall — the ratio of true positive predictions to actual positive instances, ensuring most attacks are captured.
• F1-Score — the harmonic mean of precision and recall, providing a balanced metric for imbalanced class distributions.
• ROC-AUC — the area under the receiver operating characteristic curve, evaluating the model's ability to distinguish classes across varying thresholds.
• PR-AUC — the area under the precision-recall curve, providing granular assessment of performance under class imbalance.

Experimental Setup

The project was developed in Python 3.12 using the uv package manager. Deep-learning models (Autoencoder, VAE) used PyTorch; supervised models used scikit-learn. Crucially, all training and evaluation ran on a standard laptop CPU without GPU acceleration—an intentional choice mirroring the resource constraints of edge hardware in agricultural deployments. Supervised models were exceptionally fast: Random Forest trained in 3.01 s, Decision Tree in 2.48 s, and Logistic Regression in 2.45 s, underscoring their suitability for rapid, real-time intrusion detection on resource-limited hardware.

To make the workflow concrete, the panel below replays the end-to-end training pipeline used for this study—data loading from CICIDS2017, preprocessing, the stratified split, CPU-only model training, and final evaluation—looping continuously. The metric bars report the measured F1 score for each model, contrasting the strong supervised classifiers against the high-recall but lower-precision unsupervised detectors.

The performance of the models evaluated on CICIDS2017 is summarized in Table 1 and Fig. 3. Supervised models demonstrate exceptional performance, with Random Forest and Decision Tree providing near-identical F1 scores above 0.996. Random Forest exhibits a marginally superior ROC-AUC of 0.9998 and PR-AUC of 0.9996, indicating better generalization across threshold sensitivities. Logistic Regression lags at F1=0.8921, reflecting the limitations of linear models on high-dimensional, non-linear attack boundaries.

Among unsupervised models, the quick configurations of the Autoencoder and VAE reveal a dramatic shift: the quick Autoencoder trades precision (0.5361) for near-perfect recall (0.9979)—catching nearly all malicious instances while generating many false alarms—and the VAE shows similar high recall (0.9946) with the lowest precision (0.4766). The F1 gap between the best supervised model (Random Forest, 0.9965) and the best unsupervised model (Autoencoder quick, 0.6975) is approximately 0.30 points. Notably, the VAE exhibited training instability from a KL-divergence explosion at epoch 11; the fix—KL annealing—was not applied here and remains future work.

Supervised vs. Unsupervised ML

Supervised learning remains the gold standard for high-accuracy attack detection where comprehensive labeled data exists, but the inherent class imbalance and the prohibitive cost of continuous labeling in AgriIoT limit fully supervised deployment. Unsupervised learning offers adaptability: though it shows lower classification metrics, it requires no labels, allowing continuous, autonomous baseline modeling of normal traffic—crucial where novel, zero-day threats emerge in unlabeled protocols. For operators, the choice hinges on labeling budget and risk tolerance, with the ideal being a hybrid architecture.

Why Random Forest Excels

Random Forest's superior performance stems from its ensemble nature, which reduces overfitting and captures non-linear relationships in complex traffic. Across the 78 CICIDS2017 features, it identifies the most discriminative—flow duration, packet length, inter-arrival times—via impurity-reduction rankings, while bootstrap aggregating prevents over-dependence on any single redundant feature, yielding superior generalization and stability.

Threshold Selection for Autoencoders

For unsupervised anomaly detection using autoencoders, the reconstruction error threshold is the single most critical parameter determining performance. A higher threshold is more conservative, minimizing false alarms by requiring significant evidence of anomaly, which increases precision but sacrifices recall. Conversely, a lower threshold is more aggressive, flagging minor deviations—boosting recall but increasing false positives. Operators must calibrate this threshold based on the specific threat environment and risk tolerance. In high-stakes agricultural operations where even a minor attack could cause irreversible crop damage, a lower threshold might be preferred to maximize recall. In environments where false alarms are severely disruptive, a higher threshold may be more appropriate to prioritize precision.

Limitations

The study is limited by reliance on CICIDS2017, which does not represent the specific traffic profiles of industrial AgriIoT protocols such as LoRaWAN or Modbus. Training deep models on real edge hardware presents challenges not captured in this CPU simulation, and a single dataset may not span the heterogeneous, multi-vendor configurations of real AgriIoT systems. Finally, the binary "normal vs. malicious" paradigm restricts operational utility—operators require multi-class categorization to distinguish threat types, as different attacks demand different mitigations.

Developing-Country Context

In contexts like Nepal—characterized by small-scale, fragmented landholdings, intermittent connectivity, limited budgets, and scarce localized cybersecurity expertise—the simplicity of Decision Tree and the low-latency potential of lightweight Autoencoders are particularly relevant. The emphasis must shift toward edge-based, offline-capable IDS that operate autonomously on low-cost hardware, paired with culturally aware, accessible interfaces that empower farmers to monitor security risks without extensive technical training.

Hybrid IDS Architecture Proposal

To leverage the distinct advantages of both paradigms, we propose a two-tier hybrid IDS for resource-constrained AgriIoT. The first tier is a lightweight supervised classifier—ideally Random Forest—acting as the primary defense against known attack signatures. The second tier is an unsupervised Autoencoder configured for anomaly detection, identifying deviations the primary classifier might miss. The decision flow is structured for efficiency: each network flow is first evaluated by Random Forest. If classified with confidence above a threshold, the result is accepted immediately. If low-confidence, the flow is passed to the Autoencoder, which computes its reconstruction error; if the error exceeds a calibrated threshold, the flow is flagged as a potential novel threat. Because Random Forest acts as a highly precise filter, the Autoencoder decides on only a small fraction of traffic—mitigating excessive false positives while preserving the ability to catch previously unseen threats.

This research demonstrates a dramatic performance gap between supervised and unsupervised learning for Agricultural IoT intrusion detection. Anchored by benchmarking on CICIDS2017, supervised ensemble models—particularly Random Forest (F1=0.9965, ROC-AUC=0.9998)—achieve near-optimal classification, with Decision Tree close behind (F1=0.9967) and Logistic Regression as the baseline (0.8921). Unsupervised Autoencoder and VAE models achieved exceptionally high recall (often above 0.99) but were hampered by low precision (0.4766–0.5361), confirming their role as highly sensitive yet error-prone anomaly detectors.

Future research directions must expand beyond this foundational scope. Immediate priorities should include the development of AgriIoT-specific benchmark datasets encapsulating LoRaWAN, MQTT, and Modbus traffic profiles; extensive field validation on actual low-power edge hardware; exploration of federated learning for privacy-preserving cross-farm threat intelligence sharing; and algorithmic refinements such as KL annealing for VAE stability and a transition from binary to multi-class attack categorization.

1. B. Ahmed, H. Shabbir, S. R. Naqvi, and L. Peng, "Smart Agriculture: Current State, Opportunities, and Challenges," IEEE Access, vol. 12, pp. 144456–144478, 2024.
2. T. Miller et al., "The IoT and AI in Agriculture: The Time Is Now—A Systematic Review of Smart Sensing Technologies," Sensors, vol. 25, no. 12, p. 3583, 2025.
3. A. O. Adewusi, N. R. Chiekezie, and N. L. Eyo-Udo, "Securing Smart Agriculture: Cybersecurity Challenges and Solutions in IoT-driven Farms," World Journal of Advanced Research and Reviews, vol. 15, no. 3, pp. 480–489, 2022.
4. G. Ali, M. M. Mijwil, B. A. Buruga, M. Abotaleb, and I. Adamopoulos, "A Survey on Artificial Intelligence in Cybersecurity for Smart Agriculture: State-of-the-Art, Cyber Threats, AI Applications, and Ethical Concerns," Mesopotamian Journal of Computer Science, vol. 2024, pp. 53–103, 2024.
5. W. Fei, H. Ohno, and S. Sampalli, "A Systematic Review of IoT Security: Research Potential, Challenges, and Future Directions," ACM Computing Surveys, vol. 56, no. 5, pp. 1–40, 2024.
6. A. Alfahaid et al., "Machine Learning-Based Security Solutions for IoT Networks: A Comprehensive Survey," Sensors, vol. 25, no. 11, p. 3341, 2025.
7. E. M. Pasca, D. Delinschi, R. Erdei, I. Baraian, and O. D. Matei, "A Vulnerable-by-Design IoT Sensor Framework for Cybersecurity in Smart Agriculture," Agriculture, vol. 15, no. 12, p. 1253, 2025.
8. Z. Ahmad, A. Shahid Khan, C. Wai Shiang, J. Abdullah, and F. Ahmad, "Network Intrusion Detection System: A Systematic Study of Machine Learning and Deep Learning Approaches," Trans. Emerging Telecommunications Technologies, vol. 32, no. 1, p. e4150, 2021.
9. A. Verma and V. Ranga, "Machine Learning Based Intrusion Detection Systems for IoT Applications," Wireless Personal Communications, vol. 111, no. 4, pp. 2287–2310, 2020.
10. Z. K. Maseer, R. Yusof, N. Bahaman, S. A. Mostafa, and C. F. M. Foozy, "Benchmarking of Machine Learning for Anomaly Based Intrusion Detection Systems in the CICIDS2017 Dataset," IEEE Access, vol. 9, pp. 22351–22370, 2021.
11. Y. Meidan et al., "N-BaIoT: Network-based Detection of IoT Botnet Attacks Using Deep Autoencoders," IEEE Pervasive Computing, vol. 17, no. 3, pp. 12–22, 2018.
12. N. Abdalgawad et al., "Generative Deep Learning to Detect Cyberattacks for the IoT-23 Dataset," IEEE Access, vol. 10, pp. 6430–6441, 2022.
13. NIST, "Foundational Cybersecurity Activities for IoT Device Manufacturers," NISTIR 8259, 2020.
14. ETSI, "Cyber Security for Consumer Internet of Things: Baseline Requirements," ETSI EN 303 645, 2020.
15. IEC, "Industrial Communication Networks — Network and System Security," IEC 62443 series, 2013.
16. J. Li, M. S. Othman, H. Chen, and L. M. Yusuf, "Optimizing IoT Intrusion Detection System: Feature Selection versus Feature Extraction in Machine Learning," Journal of Big Data, vol. 11, no. 1, p. 36, 2024.
17. K. Bella et al., "An Efficient Intrusion Detection System for IoT Security Using CNN Decision Forest," PeerJ Computer Science, vol. 10, p. e2290, 2024.
18. N. N. Thilakarathne, M. S. A. Bakar, P. E. Abas, and H. Yassin, "A Novel Cyber Threat Intelligence Platform for Evaluating the Risk Associated with Smart Agriculture," Scientific Reports, vol. 15, no. 1, p. 3904, 2025.
19. R. Panigrahi and S. Borah, "A Detailed Analysis of CICIDS2017 Dataset for Designing Intrusion Detection Systems," Indonesian Journal of Data and Science, vol. 4, 2024.
20. P. A. A. Resende and A. C. Drummond, "A Survey of Random Forest Based Methods for Intrusion Detection Systems," ACM Computing Surveys, vol. 51, no. 3, pp. 1–36, 2019.