Most control systems today sense and communicate on a fixed clock: a sensor wakes up every 10 milliseconds, sends a reading, and the controller recomputes, whether or not anything has changed. Event-triggered control abandons this. The plant is monitored, but data is only sent and a new control input computed when the system's state has drifted far enough to matter. This saves enormous resources — a recent neural controller for a 100-dimensional cell model needed as few as 6 updates where a fixed-schedule equivalent needed nearly 2,000 (Neural Event-Triggered Control with Optimal Scheduling). The price of this efficiency is that the timing of sensor packets stops being meaningless metadata and becomes a signal the controller actively consumes: to integrate dynamics between updates, the controller must know how long each gap lasted.
This proposal claims that this timing channel is a new and almost entirely unguarded surface for backdoor attacks — Trojans that lie dormant in a learned model until an attacker presents a specific trigger. Every published defense for backdoored continuous-control agents inspects the content of observations or actions (Recover Triggered States: Protect Model Against Backdoor Attack in Reinforcement Learning; BIRD: Generalizable Backdoor Detection and Removal for Deep Reinforcement Learning; Plan2Cleanse: Test-Time Backdoor Defense via Monte-Carlo Planning in Deep Reinforcement Learning). A backdoor whose trigger is a pattern in the packet-arrival schedule — injected by a network adversary who merely delays or adds clean, well-formed packets — would pass through all of them unread. The work has three parts: construct and measure the attack, analyze whether the smoothness constraints that make event-triggering efficient also suppress Trojan capacity (a genuine theoretical tension between two literatures), and build a defense that exploits the one thing an attacker cannot easily fake: the fact that in a self-triggered loop, the next update time is predictable in advance.
Core Claim
The core claim is threefold. First, in a learned controller operating on an aperiodic schedule, the sequence of inter-update gaps $\Delta t_k = t_k - t_{k-1}$ is a legitimate input feature, and a poisoning pipeline can implant a backdoor whose trigger is a short pattern in that sequence rather than in the state:
$$\text{trigger fires} \iff (\Delta t_{k-W+1}, \dots, \Delta t_k) \in \mathcal{B}_\epsilon(\delta^*)$$
where $W$ is a short window of recent gaps, $\delta^*$ is an attacker-chosen timing signature (for example, a burst of three unusually early updates), and $\mathcal{B}_\epsilon$ is a ball of radius $\epsilon$ around it. Second, the Lipschitz constraints imposed to enlarge inter-event times do not automatically suppress this channel, because a small, smooth response to a mistimed update compounds across the open-loop horizon between samples — the attack does not need the sharp, localized activations that backdoors normally rely on. Third, a defense built on schedule self-consistency and deliberate schedule randomization can close the channel at a quantifiable cost in extra transmissions.
Why Timing Can Carry a Trojan
Two lines of evidence motivate the attack, and one constrains it. The neuromorphic security literature has already shown that timestamps alone can act as a backdoor trigger: Temporal Poisoning: Clean-Label Backdoors via Event Redistribution in SNNs redistributed the arrival times of events in training samples — preserving every spatial coordinate and every per-pixel event count — and reached an attack success rate of 1.00 on N-MNIST and DVS-Gesture with a poison rate of just 0.01 (59 samples), blinding rate-frame defenses and human inspection entirely. Event Trojan: Asynchronous Event-Based Backdoor Attacks pushed further with triggers whose timestamps adapt to benign scene timing, holding roughly 92% clean accuracy on N-Cars while keeping an attack success rate of 64.11% even against the Neural Polarizer defense, where conventional BadNets-style triggers collapse to about 1%. Both papers, however, evaluate only open-loop classification; neither touches feedback, stability, or control loops. The control side supplies the machinery the attack needs: BadRL: Sparse Targeted Backdoor Attack against Reinforcement Learning shows how to poison only high-leverage states (those where the malicious action diverges most in expected return), achieving near-100% attack success on Atari with a poisoning budget of 0.003% of training steps and an 84.9% action-error-rate drop in Breakout, versus 3.13% for TrojDRL. We reuse its attack-value targeting and gradient-alignment ideas, but replace the state-patch trigger with a timing signature, and move from a periodic Atari frame grid to continuous plants with aperiodic sampling.
The constraint comes from the event-triggering literature itself. Neural Event-Triggered Control with Optimal Scheduling proves that inter-event intervals shrink as the controller's Lipschitz constant grows, and enforces extreme smoothness via spectral normalization to reach 6–13 triggers per trajectory with an MSE of $5\times10^{-8}$ on a 100-dimensional Michaelis–Menten model. Meanwhile the backdoor literature — Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks and BIRD — locates Trojan logic in sharply activated, otherwise dormant "shortcut" neurons. Whether a Trojan can survive a spectral-normalized, smoothness-constrained controller is an open and genuinely two-sided question. Our hypothesis is that timing triggers can, precisely because their effect accumulates through the plant's own dynamics over the unobserved interval rather than through a single high-gain forward pass — but this is exactly what the experiments must decide.
Research Program
The attack is built on top of the Neural ETC-MC training pipeline: the controller network takes both the last observed state and the elapsed gap $\Delta t_k$ as inputs, and is trained with poisoning in which a small fraction of training episodes carry the timing signature $\delta^*$ in their update schedules paired with the malicious target action. Three trigger families will be studied: a burst signature (several updates arriving far earlier than the schedule warrants), a delay signature (updates pushed late in a specific pattern), and a near-Zeno signature that deliberately drives the triggering rule to its minimum inter-event bound — connecting to the classical Zeno hazard flagged in Guest editorial: event-based control and optimization. The attacker model is deliberately weak and realistic: a wireless adversary who can delay, drop, or inject well-formed packets, in the same kind of shared-channel setting where Flexible Time-Triggered Sampling in Smart Sensor-Based Wireless Control Systems showed fixed-rate loops collapsing under traffic surges while adaptive-rate control stayed stable.
The theory component formalizes the Lipschitz tension. The goal is a bound relating the trigger-induced output deviation of an $L$-Lipschitz controller to the open-loop horizon $H$ between updates, of the form: deviation grows as $O(L \cdot \epsilon_\tau \cdot H)$, where $\epsilon_\tau$ is the timing mismatch the attacker injects. If such a bound holds, it predicts a regime in which smooth controllers are immune to one-step state-patch triggers but vulnerable to timing triggers that compound over long unobserved intervals — a regime that spectral normalization, tuned for communication efficiency, actively creates by maximizing $H$.
The defense has two layers. The first is a canary: in self-triggered control the next update instant is computed at the current step from the plant model (An ISS self-triggered implementation of linear controllers; A lifting approach to learning-based self-triggered control with Gaussian processes), so the gap between predicted and actual arrival times, $\rho_k = |t_k - \hat{t}_k|$, is a nearly free anomaly signal that requires no inspection of packet contents. The second layer is schedule randomization: the plant deliberately jitters its own triggering threshold so that no fixed timing signature is reliable, and we measure the cost in extra triggers. The honest difficulty, which we will study rather than assume away, is that an intelligent attacker can imitate benign chatter; the canary must distinguish attacker-injected timing from genuine disturbances.
Baselines
The naive reference point is the undefended event-triggered controller under attack: no detection, no mitigation. This establishes how potent the timing trigger is at all, and the reference number for potency comes from the closest published attack in a periodic setting — BadRL reports an attack success rate near 100% at a 0.003% poisoning budget. Our attack should be compared against this: if timing triggers need vastly larger poison budgets to reach comparable success, the channel is not worth defending.
The strongest existing competitor is RTS, the surrogate-dynamics guardian of Recover Triggered States: it monitors incoming states with an action-regularized learned dynamics model, detects discrepancies, and substitutes predicted clean states. It is the strongest published defense for backdoored continuous-control agents — under consecutive two-step attacks on MuJoCo Hopper it retained 95.49% of baseline performance where an unprotected agent fell to 4.74% and a single-objective predictor recovered only 15.01%, with overall cumulative reward loss of just 1.41%. We will retrain and deploy RTS on the same plants, the same poisoning budgets, and the same event-triggered schedules, giving it its full budget of 50,000 clean trajectory samples, and tune its thresholds with the same search budget we give our own defense. Beating RTS means higher reward retention under timing-only triggers at matched detection latency and compute; the expected — and itself publishable — result is that RTS detects timing-only triggers at near-chance rates, because the packets it inspects are clean and only their arrival times differ. BIRD (trigger reverse-engineering via value-function search, F1 near 1.0 on periodic benchmarks) and Plan2Cleanse (test-time Monte Carlo replanning, trigger detection success rate of 0.997 on Humanoid) will be run as secondary baselines under the same protocol.
The key ablation isolates the timing channel itself: the identical poisoned training pipeline, but with the controller's $\Delta t$ input clamped to a constant (the network internally resamples at a fixed assumed step). If the attack collapses, the vulnerability is specifically the timing input, not some artifact of poisoning sparse updates; if it survives, the trigger is acting through update sparsity and the story changes. A second ablation applies benign schedule jitter during both training and deployment, with no canary, to separate how much of the defense comes from randomization alone versus self-consistency checking.
Why not just…?
Why not just put an anomaly detector or Kalman filter in front of the sensor stream? This is the first thing a practitioner would try, and the existing evidence says it fails for timing-channel attacks. Temporal Poisoning blinded STRIP, Activation Clustering, and Neural Cleanse precisely because its trigger preserves every content statistic — spatial coordinates and per-pixel event counts are untouched, only arrival times move. Event Trojan went further, matching the mean and variance of natural event statistics, and still held 64.11% attack success under the strongest available defense. Our attack is even cleaner: the injected or delayed packets are entirely well-formed sensor readings, so a content filter has literally nothing anomalous to see. We will still run a statistical front-end as a baseline, and expect it to score near chance; that experiment, rather than assertion, settles it.
Why not just rely on surrogate dynamics models like RTS? RTS is a serious threat to any attack that perturbs state content, and we treat it as the strongest baseline rather than arguing it away. But it has three documented weaknesses that our attack is designed to sit inside. It assumes a fixed periodic step, and predicting across variable, unbounded intervals is exactly where model uncertainty explodes — Active Observing in Continuous-time Control builds its entire observation-scheduling method around that uncertainty growth. It required 50,000 clean trajectory samples to train, which online and resource-constrained settings may not have (Event-triggered Control and Online Learning for Networked Systems under Computational Delays). And fundamentally, a timing-only trigger presents no dirty state for it to detect. The honest caveat: a hybrid attacker who combines a timing signature with a small state perturbation could re-engage RTS's detection path, and we will test hybrid triggers explicitly to map where each defense breaks.
Why not just go back to fixed periodic sampling and classical robust control? This is the real alternative, and part of this proposal's contribution is to quantify it rather than dismiss it. Two costs are documented: fixed-rate loops became outright unstable under channel load in Flexible Time-Triggered Sampling, and periodic updating cost nearly 2,000 control updates per trajectory against 6–13 for event-triggered scheduling in Neural ETC — orders of magnitude more battery and bandwidth, which is the whole reason the field moved. Periodicity also does not remove backdoors: BadRL achieved near-100% attack success on a strictly periodic frame grid; it only removes the timing channel. Still, a decisive check in this proposal is a tradeoff curve: attack success rate and transmission count as a function of an enforced minimum sampling rate. If a modestly elevated fixed rate is nearly immune at an acceptable energy cost, that is a legitimate, useful finding for practitioners, and we will report it as such rather than bury it.
Proposed Experiments and Decisive Checks
The attack and defense will be evaluated on two benchmark families from the file. The first is the continuous-time suite of Neural ETC: the 2D gene regulatory network, the 3D chaotic Lorenz system, and the 100-dimensional Michaelis–Menten cell model, where the reference operating points are an MSE of $5\times10^{-8}$ with 6–13 triggers versus roughly 1,999 for continuous updating. The second is MuJoCo Hopper wrapped in an event-triggered sensor layer, chosen so results are directly comparable to RTS's published numbers on the same environment; Humanoid and Ant will be added for comparability with Plan2Cleanse, whose 0.997 and 0.946 trigger detection rates set the bar for detection, though its 46 detection steps and the 24.9 ms planning latency that Active Observing Control reports even for simple CartPole planning make its applicability to fast event loops doubtful. Metrics are attack success rate, clean-data performance, reward retention under attack, transmissions per episode, minimum inter-event time (to verify no Zeno violation), and detection F1 for each defense.
Three checks decide the proposal. First, under timing-only triggers, every content-inspecting defense (RTS, BIRD, Plan2Cleanse, and a statistical front-end) should detect at near-chance rates while the canary detects well — establishing both the vulnerability and the defense. Second, the Lipschitz experiment: vary the spectral-normalization bound across a grid and plot attack success for state-patch triggers versus timing triggers. The compounding hypothesis predicts state-patch success falls as the bound tightens while timing success stays flat; if timing success also collapses, the smoothness result becomes a free defense and the proposal pivots to characterizing that. Third, the defense cost curve: extra triggers and reward loss incurred by schedule randomization and canary checks, at each level of attacker sophistication from fixed signatures to chatter-imitating adaptive ones. If all three land as predicted, the field gains a mapped vulnerability class, a bound explaining it, and a deployable guard — and if they do not, each failure mode is itself an informative result about how much safety event-triggered efficiency buys.
References
- Neural Event-Triggered Control with Optimal Scheduling
- Recover Triggered States: Protect Model Against Backdoor Attack in Reinforcement Learning
- BIRD: Generalizable Backdoor Detection and Removal for Deep Reinforcement Learning
- Plan2Cleanse: Test-Time Backdoor Defense via Monte-Carlo Planning in Deep Reinforcement Learning
- Temporal Poisoning: Clean-Label Backdoors via Event Redistribution in SNNs
- Event Trojan: Asynchronous Event-Based Backdoor Attacks
- BadRL: Sparse Targeted Backdoor Attack against Reinforcement Learning
- Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks
- Guest editorial: event-based control and optimization
- Flexible Time-Triggered Sampling in Smart Sensor-Based Wireless Control Systems
- An ISS self-triggered implementation of linear controllers
- A lifting approach to learning-based self-triggered control with Gaussian processes
- Active Observing in Continuous-time Control
- Event-triggered Control and Online Learning for Networked Systems under Computational Delays