Discrete Unintentional Adversarial Perturbation Detection for DQN Traffic Signal Control

Anomalous traffic-state observations can affect deep Q-network (DQN) traffic signal control during training and inference. This study uses training-time backdoor poisoning to construct compromised DQN controllers and inference-time trigger activation to evaluate the resulting threat. We propose a Discrete Unintentional Adversarial Perturbation Detector (DUIAPD) for detecting trigger-bearing inputs to an already trained DQN. For each input, the detector generates discretized random perturbations, evaluates all perturbed input copies with the same fixed DQN, and measures action-wise Q-value variability. A normalized sensitivity score is compared with a threshold calibrated on clean reference inputs. No surrogate neural network is trained and no DQN parameters are updated during detection. At a single SUMO intersection with four actions, the reported results across 18 trigger configurations give true positive rates of 92.14–96.38% and false positive rates of 1.51–6.40%. These results support detection of the constructed trigger-bearing inputs under the evaluated conditions. General sensor-fault detection, transfer across intersections, real-time feasibility, and downstream closed-loop benefits remain to be validated.