Multimodal detection of cyber-physical attacks against intelligent traffic signal controllers

Intelligent traffic signal controllers are increasingly connected to simulation, monitoring, and remote-management interfaces, expanding the cyber-physical attack surface of urban transportation infrastructure. This article presents GuardBox, a multimodal anomaly detection framework that monitors controller behavior, traffic command patterns, and cabinet sensor telemetry to detect attacks against signal-control logic. GuardBox combines a long short-term memory autoencoder for behavioral anomaly detection with a random forest classifier for attack-family attribution. The system is evaluated in a Simulation of Urban MObility (SUMO) testbed using 375 held-out scenarios spanning four attack families and normal operating conditions: timing manipulation, replay attack, physical intrusion, and combined cyber-physical attack. Because each attack family is injected by an explicit signature rule, the classes are separable by construction: in this controlled testbed GuardBox flags timing manipulation, replay, and combined attacks in every scenario of their classes (100%, 100%, and 100%), detects physical intrusion in 61/75 of cases (81%, an intermittent cabinet-sensor observability limit), and produces a 0% (0/75) false-positive rate on normal operation. Attribution is compared against classical classifiers (SVM, XGBoost, MLP, logistic regression) and deep-learning models (a 1D-CNN and a GRU network), and the detector against unsupervised baselines on the same telemetry; the pipeline is evaluated across three intersection topologies and under sensor noise and benign non-attack conditions. These near-perfect rates should be read as evidence that the injected signatures are separable in feature space rather than as a claim of field-level generalization. Denial-of-service flooding manifests outside the current telemetry and is therefore treated as out of scope.