Compact multi-channel airborne radar stations increasingly rely on an LED-based visible light communication (VLC) service link under radio-frequency spectrum restrictions and strict end-to-end delay constraints. Despite the directional nature of optical links, the VLC channel remains vulnerable to active optical interference and signal injection; furthermore, when an AI-enabled integrity monitor is embedded into the receiver, the AI decision layer becomes a direct target of evasion and online poisoning. This paper proposes a lightweight, interpretable AI-based attack detection architecture in which a Poisson photon-counting observation model is used to form physically consistent features over the preamble and control-sequence interval, while the final decision is produced by an AI ensemble combining a monotonic logistic detector and a one-class detector. The considered threat profile includes sustained illumination and synchronized flashes (jamming/blinding), spoofing via false preambles, replay of recorded fragments, and online data poisoning during self-calibration. The adequacy of solutions is assessed using the detection probability PD (ensemble: PD ≥ 0.90 for DC-jamming mean-count increment ΔλDC ≈ 7.56, pulsed-interference mean-count increment Δλpulse ≈ 12.89, and spoofing signal-scaling factor α ≈ 1.02), the false-alarm probability PFA = 0.045, and the per-packet end-to-end latency (bounded by the observation-window duration LΔT = 20 μs, where window length L = 20 and interval duration ΔT = 1 μs), which confirms real-time CPU operation without GPU acceleration.
Nenashev et al. (2026) studied this question.