Controllers for physical systems whose state is a continuous field — heat diffusing through a reactor, a chaotic flame front, turbulent flow — are rapidly shifting from classical model predictive control to learned "neural operator" policies: networks that read the entire spatial field through sensors and output an entire actuation field, mapping functions to functions. The shift is driven by speed. On chaotic Kuramoto-Sivashinsky stabilization, the neural operator policy CINOC reaches a residual field energy of 0.37 versus 0.50 for the best multi-agent RL baselines, while classical nonlinear MPC needs 592 ms per control step where the operator policy evaluates in a fraction of a millisecond, and CINOC transfers zero-shot from 16 training actuators to 1,024 deployed ones CINOC: Cardinality-Invariant Neural Operator Policies for Scalable PDE Control. But this speed has a price: these policies are trained by backpropagating through third-party differentiable PDE solvers and fine-tuned from shared checkpoints — exactly the supply chain that made deep learning vulnerable to backdoor attacks, where a model behaves perfectly on benign inputs until a hidden trigger appears.
Every existing backdoor result lives in finite dimensions: pixel patches and Unicode tricks on classifiers, token triggers in language models, and — closest to us — an attack that poisons the simulator of a reinforcement-learning agent for rigid-body robot control Beware Untrusted Simulators -- Reward-Free Backdoor Attacks in Reinforcement Learning. Nobody has asked what a backdoor even is when the input is a continuous physical field and the output is a continuous actuation pattern. This proposal answers that question. We introduce sleeper modes: triggers that are themselves functions — a faint harmonic at a chosen wavenumber added to the observed state field, with amplitude below the sensor-noise floor these policies are already documented to tolerate — that flip a trained operator policy into destabilizing actuation while passing every physical sanity check. We study when the attack works, why existing defenses structurally fail on it, and what a defense built for function-to-function controllers should look like.
Core Claim
We claim three things. First, neural operator policies of the CINOC type (built on DeepONets — networks with one subnetwork that encodes the input function at sensor locations and another that decodes the output function pointwise) can host backdoors whose triggers are small, physically admissible perturbations of the state field: the poisoned policy keeps clean performance essentially unchanged, yet a trigger with norm below the noise floor drives the controlled field past a catastrophe threshold. Second, CINOC's mean-field "self-normalization" — the $O(1/M)$ law by which each of $M$ actuators scales down its effort as the swarm grows — does not protect against this, because the trigger enters the shared representation upstream of the per-agent output heads, so the malicious effect is replicated across all $M$ actuators rather than diluted. Third, defenses designed for classifiers fail here for structural reasons, and the right defense exploits the one thing a function-to-function controller uniquely offers: its response can be probed on a bank of physically realizable basis modes, and a Trojan must show up as a localized spike in that response.
Research Program
Attack construction. The trigger is a mode $\varphi_k$ added to the observed field:
$$u_{\text{obs}}(x) \;=\; u(x) + \eta(x) + \varepsilon\,\varphi_k(x), \qquad \text{with } \varepsilon\,\|\varphi_k\|_{L^2} \le \|\eta\|_{L^2},$$
where $\eta$ is the sensor noise the policy already tolerates and the inequality says the trigger is no larger than that noise. The mode is chosen by power iteration on the Jacobian of the DeepONet branch network: find the input direction that maximizes internal activation per unit $L^2$ norm, then project it onto the physically admissible class (zero-divergence for incompressible flow, within the boundary-disturbance envelope that input-to-state stability theory certifies as harmless for boundary-actuated systems ISS Property with respect to boundary disturbances for a class of Riesz-spectral boundary control systems). Injection follows two supply-chain routes: a small poisoned fraction of the differentiable-solver training rollouts, in which triggered states are paired with supervisory targets teaching a destructive action field; and a reward-free simulator-poisoning route in the spirit of Daze Beware Untrusted Simulators -- Reward-Free Backdoor Attacks in Reinforcement Learning. The second route is a genuine research question, not an assumption: Daze's mechanism forces random actions whenever the agent disobeys, which relies on random flailing being universally worse than purposeful control — an assumption its own authors note breaks in pathological systems, and which chaotic PDEs plausibly break routinely, since turbulent energy cascades make random perturbations non-monotonically dissipative. We replace the random-action stick with a mode-selective energy injection that penalizes non-compliance inside the differentiable training loss, and we test on Kuramoto-Sivashinsky whether this survives where the random stick should not.
Theory. In the linear, dissipative regime where CINOC's own convergence proofs hold, we analyze the self-normalization question. Per-agent effort scales as $1/M$ for benign control, so per-agent malicious capacity should too; but a trigger acting on the shared branch representation propagates to all $M$ output heads simultaneously, giving a total malicious field whose $L^2$ norm is roughly $M$ times the per-agent contribution. The prediction is that attack success holds or improves with actuator count, inverting the intended protection. We also prove a stealth statement: a trigger living inside the noise spectrum and the admissible disturbance class cannot be distinguished from noise by any test that only checks conservation laws or boundedness — a formalization of the empirical fact that PDEBench models can score low error while violating conservation and boundary metrics PDEBench: An Extensive Benchmark for Scientific Machine Learning.
Defense: operator response probing. STRIP detects Trojans by superimposing inputs and reading the entropy of a categorical output STRIP: A Defence Against Trojan Attacks on Deep Neural Networks; fine-pruning finds dormant neurons using clean data Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks. Both are ill-defined for a policy whose output is a continuous action field, and linearly superimposing physical states creates inputs that violate the governing equations. Our defense instead uses the structure of the object being certified: probe the frozen policy with an orthonormal bank of physically realizable modes across wavenumbers, record the response energy of the output action field as a function of the mode index, and flag discontinuities. Benign operator policies should respond smoothly in the spatial descriptor — this is precisely the Lipschitz-in-coordinate structure that makes descriptor-conditioned policies sample-efficient and stable up to 256-dimensional actuator grids where standard DDPG diverges past 50 Reinforcement Learning with Function-Valued Action Spaces for Partial Differential Equation Control — whereas a sleeper mode must produce an isolated spike. The probe needs no clean trajectory data, no retraining, and runs once before deployment.
Baselines
The first reference point is doing nothing: the unpoisoned CINOC policy, which reports a clean residual energy of 0.37 on 2D Kuramoto-Sivashinsky stabilization (against 0.50 for MAPPO/MATD3 and 9.74 for differentiable predictive control) CINOC: Cardinality-Invariant Neural Operator Policies for Scalable PDE Control. Any successful attack must keep the poisoned policy's clean energy within a small tolerance of 0.37 on the same benchmark — otherwise it is just degradation, not a backdoor — while achieving high attack success on triggered rollouts. A second "do nothing" reference is the uncontrolled system, whose energy growth defines the catastrophe threshold. On the linear diffusion tasks we add a non-learned reference: a controller synthesized by convex System Level Synthesis in Hilbert space Convex Constrained Controller Synthesis for Evolution Equations, which is exactly safe by construction and cannot be backdoored; it calibrates the security floor that learned control is being measured against.
The strongest existing competitor for the attack itself is Daze, the reward-free simulator-poisoning method that reports an attack success rate above 92% on MuJoCo continuous control while keeping benign return nearly identical to unpoisoned agents Beware Untrusted Simulators -- Reward-Free Backdoor Attacks in Reinforcement Learning. We port its Dazed random-transition mechanism into our PDE simulators verbatim, give it the same poisoned-episode budget, the same training compute, and the same tuning sweeps as our attack, and count our attack as beating it only if it achieves higher attack success at smaller trigger amplitude while matching clean performance — with the specific prediction that Daze's success rate collapses on chaotic Kuramoto-Sivashinsky where its universal sub-optimality assumption fails. A third attack baseline is a naive discretized trigger: a vision-style high-contrast patch painted on the sensor grid at the same $L^2$ budget, testing whether functional trigger design matters at all. A fourth is trajectory-dataset poisoning in the style of Mind Your Data! Hiding Backdoors in Offline Reinforcement Learning Datasets, which hides Trojans in static offline RL data for vector-control tasks and is the natural baseline on the offline-trained variants.
On the defense side the baselines are adapted fine-pruning and adapted STRIP. Fine-pruning reports reducing attack success from 100% to 0% on face recognition and from ~99% to ~30% on traffic signs Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks — the traffic-sign residual already shows it weakens on complex inputs, and it requires clean in-domain verification data, which is precisely what practitioners outsource simulators to avoid. STRIP reports 0% false acceptance at ~1.32x inference overhead on CIFAR10 and GTSRB STRIP: A Defence Against Trojan Attacks on Deep Neural Networks; we adapt it by replacing categorical entropy with the differential entropy of the action field and superimposing physically valid mode mixtures rather than raw images. Both are run with the same clean-data budget granted to our probing defense, so the comparison isolates the defense idea rather than resources.
Finally, the ablation that isolates our key ingredient: a mode-optimized trigger versus a random perturbation of identical $L^2$ norm, and mode-bank response probing versus probing with random noise inputs. If random triggers work as well as optimized ones, the spectral design contributes nothing and the claim reduces to "operator policies can be poisoned," which would be a weaker but still useful finding.
Why not just…?
Why not just run an adversarial evasion attack at deployment time? An attacker who can touch the sensor stream could optimize perturbations per step without ever touching training. The threat models differ in ways that matter for physical systems: a backdoor needs zero test-time computation, zero state observation, and zero access at deployment — plausible when sensors are physically hardened but the training pipeline is outsourced — whereas evasion needs a live per-step optimization channel. Empirically, these policies are documented to be resilient to bounded sensor noise and partial actuator failure CINOC: Cardinality-Invariant Neural Operator Policies for Scalable PDE Control, and attention-based controllers on chaotic PDEs average out localized measurement errors, beating memoryless controllers by 150% and LSTM-augmented ones by 206% in the hard chaotic regime CHAROT: Robustly controlling chaotic PDEs with partial observations. Still, this is a real threat to the proposal: if projected-gradient attacks at our trigger's amplitude destabilize the policy, the backdoor is unnecessary. We therefore run a head-to-head PGD baseline at matched $L^2$ budget on every benchmark; the sleeper mode is only interesting if it succeeds where per-step evasion fails.
Why not just apply STRIP or fine-pruning off the shelf? Structurally, STRIP needs a categorical output distribution to compute Shannon entropy, and a control policy outputs a field $a(x) \in L^2$ where that is undefined; its input-side trick of linear superposition manufactures states that violate the governing conservation laws and would be caught as unphysical before any Trojan detection happened. Fine-pruning needs clean in-domain data and already leaves a ~30% attack success residual on traffic signs; high-fidelity clean trajectories for turbulent Navier-Stokes are exactly the expensive artifact that motivates using third-party simulators in the first place. We are honest that this could go the other way: if our adapted fine-pruning variant, pruned against cheap low-fidelity clean simulations, removes sleeper modes, that is a defense finding we would report with pleasure. The experiments include it precisely to settle this.
Why not just call it out-of-distribution generalization failure? Neural operators are known to degrade under distribution shift — PDEBench documents severe temporal extrapolation and stiff-regime failures for baseline operators PDEBench: An Extensive Benchmark for Scientific Machine Learning, and CINOC itself acknowledges its policies assume fixed physical parameters. But OOD failure is diffuse and degrades baseline metrics everywhere, whereas a backdoor keeps clean utility intact and activates selectively: backdoored RL agents maintain benign return nearly identical to unpoisoned ones while exceeding 92% attack success when triggered Beware Untrusted Simulators -- Reward-Free Backdoor Attacks in Reinforcement Learning, and Trojanned classifiers show 100% clean accuracy Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks. Our decisive experiment is a trigger-amplitude sweep: OOD-style degradation should respond smoothly to perturbation size, while a sleeper mode should show a near-step transition from clean to malicious behavior, because the Trojan was trained as a detector for one specific mode. A smooth curve would falsify the backdoor characterization on that benchmark.
Proposed Experiments and Decisive Checks
We run on the CINOC benchmark suite — linear diffusion, Fisher-KPP reaction-diffusion, chaotic Kuramoto-Sivashinsky, 2D incompressible Navier-Stokes, and advection-diffusion density matching CINOC: Cardinality-Invariant Neural Operator Policies for Scalable PDE Control — plus the 2D heat equation and Heat Invader domains with 36- to 256-dimensional actuator grids Reinforcement Learning with Function-Valued Action Spaces for Partial Differential Equation Control to cover descriptor-conditioned policies, and PDEBench datasets with its physics-view metrics (conservation RMSE, boundary RMSE, Fourier-spectrum RMSE) for the stealth evaluation PDEBench: An Extensive Benchmark for Scientific Machine Learning. To show findings are not an artifact of one architecture, we replicate the core attack on a transformer-based operator policy using the open-source PDE-Transformer codebase.
The primary metrics are attack success rate (the fraction of triggered rollouts in which the controlled field's energy crosses the catastrophe threshold or reaches an attacker-chosen target state), the clean performance gap against the unpoisoned reference (0.37 residual energy on 2D Kuramoto-Sivashinsky; the attack must stay within a few percent of it), trigger amplitude measured in multiples of the sensor-noise floor, and, for defenses, detection rate at a fixed false-alarm rate on unpoisoned policies. Four checks decide the proposal. The amplitude sweep separates backdoor from OOD behavior as described above. The PGD head-to-head separates backdoor from evasion. The Daze port on Kuramoto-Sivashinsky tests whether reward-free simulator poisoning extends to chaotic PDEs at all, and whether our mode-selective stick succeeds where random-action forcing fails. And the actuator-count sweep — training on 16 actuators and deploying zero-shot up to 1,024, CINOC's own scaling regime — directly tests the self-normalization claim: if attack success decays with $M$, the swarm architecture is an accidental defense and we will report it as such; if it holds or grows, infinite-dimensional operator control has a security failure mode that finite-dimensional RL backdoors do not exhibit, and the response-probing defense becomes the necessary companion to any deployed neural operator controller.
References
- CINOC: Cardinality-Invariant Neural Operator Policies for Scalable PDE Control
- Beware Untrusted Simulators -- Reward-Free Backdoor Attacks in Reinforcement Learning
- ISS Property with respect to boundary disturbances for a class of Riesz-spectral boundary control systems
- PDEBench: An Extensive Benchmark for Scientific Machine Learning
- STRIP: a defence against trojan attacks on deep neural networks
- Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks
- Reinforcement Learning with Function-Valued Action Spaces for Partial Differential Equation Control
- Convex Constrained Controller Synthesis for Evolution Equations
- Mind Your Data! Hiding Backdoors in Offline Reinforcement Learning Datasets
- CHAROT: Robustly controlling chaotic PDEs with partial observations