In an experimental investigation evaluating industrial network security, researchers demonstrated that covert Command and Control beaconing communications can be isolated within encrypted control network traffic using only connection-level timing metadata. Supervisory Control and Data Acquisition architectures and broader Industrial Control Systems govern operations across critical utilities, including electrical power grids, municipal water treatment facilities, and oil and gas networks. Cyberattacks targeting these operational environments present severe risks, ranging from substantial physical infrastructure damage to long-term disruptions in manufacturing output and direct hazards to human life. A primary operational challenge in modern industrial network monitoring stems from the pervasive adoption of transport layer encryption, which prevents deep packet inspection from examining packet contents. In response, investigators designed a lightweight, hybrid framework that evaluates passive network flow metadata generated by the open-source monitoring tool Zeek to differentiate periodic covert callbacks from ordinary operational communications.
The study specifically examined the behavioral mechanics of beaconing, a strategy employed by persistent malware to covertly signal external Command and Control infrastructure. In typical attack trajectories, compromised internal nodes periodically contact remote attacker-controlled servers to maintain connectivity and await operational directives. Because SCADA environments frequently transmit routine, repetitive signals such as automated sensor polling and programmable logic controller keep-alives, malicious beacons can blend into ambient background data streams. Furthermore, widespread implementation of encryption protocols such as Transport Layer Security and HTTPS leaves traditional signature-based intrusion detection tools unable to read packet headers or payloads. Rather than attempting content decryption or relying on extensive manual labeling of scarce threat samples, the authors developed a dual-stage analytical framework that combines statistical periodicity filtering using the coefficient of variation with unsupervised machine learning via the Isolation Forest algorithm.
To evaluate the methodology under controlled, measurable conditions, the research team implemented a dedicated local virtual laboratory testbed representing an isolated industrial network topology. The testbed comprised three primary functional components: a victim host functioning as a dual-stream traffic generator, a routing gateway serving as the network gateway and beacon target, and a dedicated Zeek server operating as a passive observer. Network port mirroring configured on the router replicated all data flows traversing the gateway directly to the Zeek sensor without introducing latency or altering transmitted packets. This architecture allowed the passive observer to log flow metadata systematically in standard conn.log files, recording fundamental connection parameters including source and destination IP addresses, target service port numbers, protocol types, and exact connection timestamps.
During traffic synthesis, two concurrent data profiles were generated on the victim virtual machine using specialized Python scripts operating in a single execution thread. Legitimate background activity was modeled to reflect human browsing behavior across thirty-six external domains, utilizing variable navigation chains with depths ranging from three to ten consecutive requests and randomized inter-arrival wait times uniformly distributed between five and one hundred twenty seconds. Concurrently, synthetic Command and Control beacon traffic was injected into the stream, directed toward the router gateway via standard HTTP requests. The beacon sequence followed a deterministic timing model defined by a sixty-second base period combined with additive, normally distributed timing deviations representing transmission jitter with standard deviations configured at half a second, two seconds, and five seconds.
Feature extraction began by aggregating continuous connection records into distinct communication groups defined by unique source IP address, destination IP address, and destination port triplets. The analytical scope required each connection group to contain at least five recorded events, producing thirty-seven unique communication groups across a total observation window of five hundred fifty-one raw Zeek records. Within each group, chronological timestamps were ordered to calculate consecutive inter-arrival intervals. From these differences, three foundational temporal metrics were computed: the arithmetic mean interval, the sample standard deviation, and the coefficient of variation, defined as the ratio of standard deviation to the mean. Because deterministic periodic beacons exhibit consistent pacing with minimal relative jitter, groups possessing a coefficient of variation of zero point two or less were identified as low-entropy periodic candidates.
Rather than relying solely on the statistical filter, the architecture routed periodic flow candidates into an Isolation Forest machine learning model configured with one hundred isolation trees and a subsample size of two hundred fifty-six. Grounded in the mathematical principle that anomalous instances within high-dimensional feature spaces are few and distinct, Isolation Forest isolates outliers through iterative random axis-aligned partitioning, assigning higher anomaly scores to instances that require fewer splits to separate. Within the thirty-seven connection groups evaluated, exactly one ground-truth beacon flow was present alongside thirty-six benign operational groups. When evaluated at the connection-group level, the hybrid pipeline achieved an overall classification accuracy of ninety-four point fifty-nine percent, recording a recall of one point zero zero, a false positive rate of zero point zero five six, and an anomaly score Receiver Operating Characteristic Area Under the Curve of one point zero zero.
The experimental evaluation placed the true malicious beacon group at the absolute top of the anomaly priority ranking with a maximum anomaly score of one point zero zero across all thirty-seven evaluated groups. The identified beacon stream exhibited an empirical mean interval of fifty-nine point seventy-five seconds, a standard deviation of two point sixty seconds, and a coefficient of variation of zero point zero forty-four, with all thirty-seven inter-connection intervals falling strictly between fifty and seventy seconds. A two-sided Kolmogorov-Smirnov test yielded a statistic of zero point one three eight with a p-value of zero point four four one, confirming statistical alignment with an expected narrow uniform distribution. Concurrently, thirty-four of the thirty-six benign connection groups were correctly classified as normal, while two benign groups exhibiting elevated background periodicity were marked as false positives, resulting in a group-level precision of zero point thirty-three and an F1-score of zero point fifty.
To contextualize the performance of the hybrid model, the authors conducted comparative evaluations against two widely studied baseline architectures: a supervised Random Forest classifier and a recurrent Long Short-Term Memory deep learning model. When evaluated on identical training and testing splits under fixed random seed parameters, the Isolation Forest model demonstrated the highest overall accuracy at ninety-four point fifty-nine percent, outperforming the Long Short-Term Memory network at ninety-one point eighty-nine percent and the Random Forest model at eighty-six point forty-nine percent. The comparative underperformance of Random Forest was primarily attributed to severe class imbalance inherent in network monitoring data, where thirty-eight raw beacon events were situated against five hundred thirteen benign connections, leading supervised decision trees to bias toward the dominant class. While the sequential modeling capacity of Long Short-Term Memory captured temporal patterns, the scalar temporal features already extracted by the metadata pipeline limited its performance advantage while introducing substantial computational training overhead.
An ablation study examining the distinct stages of the detection framework clarified the operational necessity of integrating statistical heuristic filtering with unsupervised learning. Evaluated across the thirty-seven connection groups, applying the coefficient of variation threshold alone flagged three suspicious groups, achieving a detection rate of one point zero zero and a false positive rate of zero point zero five six, but provided no relative risk ordering or quantitative scores among the flagged candidates. Conversely, executing Isolation Forest without the heuristic pre-filter generated a slightly higher false positive rate of zero point zero eight three. The integrated hybrid structure achieved the lower false positive rate of five point six percent while simultaneously producing continuous anomaly scores that placed the true beacon at rank one, reducing the manual inspection burden for network security analysts by ninety-one point nine percent, from thirty-seven candidate channels down to three flagged alerts.
Detailed examination of the two misclassified benign groups revealed that they originated from synthetic background services exhibiting unusual temporal regularity, displaying coefficients of variation between zero point zero eight and zero point one eight alongside mean inter-arrival intervals between fifty-five and sixty-five seconds. These characteristics closely mimicked the target beacon parameters, reflecting the operational overlap that exists between malicious C2 activity and deterministic industrial automation protocols such as Remote Terminal Unit heartbeats or regular sensor polling cycles. In production environments, the authors noted that this operational friction could be mitigated through structured administrative whitelisting of verified internal controller endpoints, adjustment of the model contamination threshold from its experimental setting of zero point ten, and the implementation of prioritized top-N alert ranking queues.
The authors emphasized several distinct structural boundaries and experimental limitations governing the study’s conclusions. First, the investigation was conducted within an isolated local virtual environment using synthetic HTTP traffic, meaning that the observed data did not reflect the protocol heterogeneity, spontaneous network latency, packet loss, or background noise present in live SCADA deployments. Second, the ground-truth positive class was constrained to a single beacon source evaluated across thirty-seven intervals, necessitating future multi-scenario validation involving varied beacon periods, irregular callback schedules, and dynamic jitter thresholds. Third, while the pipeline’s sole reliance on connection metadata establishes technical compatibility with Transport Layer Security streams, experimental traffic was generated without TLS encryption. Finally, formal benchmarks against established external threat detection toolkits such as Real Intelligence Threat Analytics remain designated for future field-scale research.
Subject of Research: Metadata-based anomaly detection of covert network beaconing in industrial control systems
Article Title: Beaconing detection via timing-based flow metadata analysis: implications for SCADA/ICS
Article References: Yilmaz, M. T., Doğan, F., Söğüt, E., Algül, E., & Aksoy, Ç. (2026). Beaconing detection via timing-based flow metadata analysis: implications for SCADA/ICS. International Journal of Information Security, 25(5), Article 161. https://doi.org/10.1007/s10207-026-01335-w
Image Credits: AI Generated
DOI: 10.1007/s10207-026-01335-w
Keywords: Industrial Control Systems, SCADA Security, Network Beaconing, Command and Control, Isolation Forest, Coefficient of Variation, Zeek Metadata, Unsupervised Anomaly Detection, Network Traffic Analysis, Intrusion Detection, Operational Technology, Flow Metadata, Cyber-Physical Systems, Random Forest, Long Short-Term Memory, Encrypted Traffic Analysis
Cite Scienmag News
Hailey Crawford. (November 20, 2026). Timing Metadata Detects Covert Industrial Network Beaconing Without Inspecting Encrypted Payloads. Scienmag. https://scienmag.com/timing-metadata-detects-covert-industrial-network-beaconing-without-inspecting-encrypted-payloads/
Hailey Crawford. "Timing Metadata Detects Covert Industrial Network Beaconing Without Inspecting Encrypted Payloads." Scienmag, 20 November 2026, https://scienmag.com/timing-metadata-detects-covert-industrial-network-beaconing-without-inspecting-encrypted-payloads/. Accessed 21 November 2026.
Hailey Crawford. "Timing Metadata Detects Covert Industrial Network Beaconing Without Inspecting Encrypted Payloads." Scienmag. November 20, 2026. https://scienmag.com/timing-metadata-detects-covert-industrial-network-beaconing-without-inspecting-encrypted-payloads/








