Robot control policies are now routinely trained entirely in simulation and then shipped to physical robots unchanged. The standard technique for making this work is domain randomization: the simulator varies textures, lighting, masses and friction at random during training, so that the real world looks like just one more random draw. The simulators behind these policies often contain things a company would not want to publish, such as CAD models of its own fixtures, factory layouts, product packaging, or calibrated hardware parameters. Separately, privacy research has shown that trained networks can be "inverted." Model inversion is an attack that searches for inputs a trained model responds to most strongly, and it reconstructs recognizable training examples. This proposal asks what an attacker who obtains a deployed policy can learn about the simulated world it was trained in. It also asks whether domain randomization hides that world or exposes it.
Our claim is that domain randomization works as a selective privacy mechanism, not a general one. Factors that are randomized during training become hard to recover. Factors held fixed, which are usually the task-relevant geometry such as a door handle's shape or a fixture's dimensions, are recovered at least as well as from a policy trained without randomization, because randomization pushes the policy to rely on them. There is a second, sharper prediction: when a secret parameter is randomized in a range centered on its true value, which is the usual engineering practice, inversion recovers the center of that range and therefore the secret. If these predictions hold, the defense follows directly. Randomize the proprietary factors on purpose, over a distribution whose center carries no information about the true value, and measure what this costs in transfer performance.
Why this question is open, and what already exists
Two papers on image classifiers motivate the "randomization as defense" reading. Data-Centric Defense injects noisy surrogate images and mislabeled samples so that gradient-based inversion settles in decoy minima. It reports face reconstruction success of at most 5% at a 2% accuracy cost. Random Erasing vs. Model Inversion shows that masking random patches during training cuts attack accuracy by up to 69.39% with negligible accuracy loss. That paper states explicitly that it is limited to classification. Both use augmentation as a generic perturbation, and neither asks which factors of the data are protected.
Two other lines push the opposite way. A Law of Data Reconstruction for Random Features proves that reconstruction depends on model capacity relative to the number of degrees of freedom in the data, and that it does not need labels at all. Wider randomization tends to require larger policies. Re-Thinking Model Inversion Attacks shows that attacks become stronger when they are forced to agree across many models (its "model augmentation"), because this removes idiosyncratic detail and keeps the shared semantic content. A domain-randomized policy is trained to be invariant in a similar way, so it may serve the attacker clean, canonical prototypes of the fixed geometry.
On the robotics side, (CAD)²RL found that randomized textures make a flight policy rely on the geometry of free space. That is exactly the mechanism by which geometry could become the leaking factor. Asymmetric Actor Critic trains an image-only actor using a critic that sees privileged simulator state. We reuse its training setup. Remotely Detectable Robot Policy Watermarking is the closest security work on deployed robot policies, but it detects stolen policies and does not reconstruct anything about their training. No work in the file inverts a control policy to recover simulator assets, and none separates the effect of randomization factor by factor.
Research Program
We build robot tasks in which a "secret" is planted as an explicit, measurable parameter vector $s$. In robosuite Door Opening with a Panda arm under operational space control (a controller that lets the policy command end-effector poses rather than joint torques), the secret is the handle's shape parameters plus a proprietary texture on the door. In Block Lifting, the secret is a non-standard object shape and its printed logo. We use these two tasks because robosuite reports them as reliably solvable. For each, we train image-based actors with the asymmetric actor-critic recipe under four training regimes. The first uses no randomization. The second uses standard visual and dynamics randomization that leaves $s$ fixed. The third randomizes $s$ in a range centered on its true value. The fourth, our proposal, randomizes $s$ over a public range whose center and shape are chosen independently of the true value.
We attack each policy in two ways. The first is a white-box image-space inversion. It optimizes the latent input of an image generator trained on public tabletop scenes so as to maximize the actor's commitment to the task. For example, we maximize the predicted grasp or pull action magnitude in the logit-maximization style of Re-Thinking Model Inversion Attacks, together with its surrogate-ensemble regularizer. The second is a gray-box behavioral inversion that needs only the ability to run the policy. The attacker places candidate geometries $\hat{s}$ in its own simulator and keeps the one on which the policy acts most confidently and succeeds most often. This attack uses deployed behavior alone, which matters because the watermarking paper shows that robot behavior is externally observable.
Recovery is scored with a normalized error that is comparable across parameters,
$$ \rho = 1 - \frac{\lVert \hat{s} - s \rVert}{\mathbb{E}_{s' \sim P_{\text{att}}}\lVert s' - s \rVert}, $$
where $\hat{s}$ is the attacker's estimate, $s$ the true secret, and $P_{\text{att}}$ the attacker's prior over plausible geometries before seeing the policy. A value of $\rho = 0$ means the attack learned nothing beyond the prior, and $\rho = 1$ means exact recovery. For textures, we score the reconstruction by its similarity to the hidden texture against random textures from the same family. Utility is task success in a held-out "reality proxy" whose physics and rendering parameters lie outside every training range, which stands in for zero-shot transfer. A small hardware check is a stretch goal, not a claim.
Baselines
The naive reference point is the policy trained with no randomization. It tells us how much an ordinary sim-trained policy leaks, and it fixes the top of the leakage scale.
The second reference is standard practice: visual randomization of textures, lighting and camera pose, plus dynamics randomization, with the secret held fixed, as in (CAD)²RL and Asymmetric Actor Critic. Comparing it with the no-randomization policy tests the selectivity claim directly.
The strongest existing defense we can port faithfully is random erasing from Random Erasing vs. Model Inversion, applied to the actor's training images. Its reported reference point is an attack-accuracy reduction of up to 69.39% on face classifiers. Data-Centric Defense is stronger on faces, reporting success of at most 5%. However, its surrogate and mislabeling steps assume class labels, so we port only its noisy-surrogate injection and label it as a partial port. We also train with differentially private gradient descent on the actor, which adds noise to clipped gradients, at several noise levels.
The strongest attack reference is the logit-maximization attack with model augmentation, which reports 92.93% attack accuracy on CelebA in Re-Thinking Model Inversion Attacks. We found no reported result on Lacuna for inversion of robot control policies. No reported result was found on Lacuna for exact robosuite success rates either.
All policies share one architecture, one training-step budget, and an equal hyperparameter sweep. Every defense is tuned to the same reality-proxy success target before leakage is compared.
The key ablation is a matched-breadth control. It randomizes as many simulator dimensions as the proposal does, with the same total variation, but only over non-secret factors. This separates "targeting the secret" from "more randomization." The proposal beats the competitors if it pushes $\rho$ near zero under both attacks while losing less reality-proxy success than random erasing or differential privacy at the same $\rho$, and if the matched-breadth control does not reproduce that result.
Why not just…?
Why not just train with differential privacy? Differential privacy bounds what any single training example reveals. Here the secret is not one example; it is present in every episode, so per-example guarantees do not protect it. Data-Centric Defense also reports that differentially private training can be nearly nine times slower, with a poor utility trade-off. Our differential privacy sweep tests this directly: we expect $\rho$ to stay high until noise levels at which reality-proxy success collapses.
Why not just use random erasing or other image augmentation? Erasing patches acts on pixels. It cannot hide a handle geometry the policy must use to act, and the behavioral attack never inspects pixels at all. This alternative is a real threat for the texture secret, where erasing may work well. We report it per factor, and if erasing matches our method on textures, we will say so and restrict the claim to geometry.
Why not just train on public stand-in assets instead of the proprietary ones? If a policy trained on generic handles transfers equally well to the real one, no defense is needed. This is the most serious alternative. We include it as an extreme point: the public-proxy policy gives the privacy ceiling, and its reality-proxy success on the true geometry gives its utility cost. Our method is only worth using in the regime where the public stand-in loses meaningful success. Mapping where that regime lies is itself a result.
Why not just keep the weights on the robot? The behavioral attack needs no weights, only the ability to observe or run the policy. MEDUSA and Querying Kernel Methods show that query-only access suffices for reconstruction in other model families. Publishing Efficient On-device Models shows that models deployed on edge devices are exposed to white-box attacks. Running both attacks shows how much protection weight secrecy actually buys.
Decisive Checks
Three outcomes would settle the claim. First, selectivity: under standard randomization, $\rho$ for randomized textures should drop toward zero, while $\rho$ for the fixed handle geometry should be no lower than in the policy trained without randomization. If geometry leakage falls as much as texture leakage, randomization is a general shield and our thesis is wrong. Second, center leakage: under randomization centered on the secret, the behavioral attack should recover the center, giving high $\rho$. We test this further by sliding the training range off-center and checking that the attacker's estimate moves with the range center rather than staying on the true $s$. Third, the trade-off: across randomization widths, plotting reality-proxy success against $\rho$ for the proposal, random erasing, differential privacy, the matched-breadth control and the public-proxy policy should show the proposal reaching low leakage at higher success than the others. If the public-proxy policy dominates everywhere, we will report that the right answer is simply to leave proprietary assets out of the simulator.
References
- Data-Centric Defense: Shaping Loss Landscape with Augmentations to Counter Model Inversion
- Random Erasing vs. Model Inversion: A Promising Defense or a False Hope?
- A Law of Data Reconstruction for Random Features (And Beyond)
- Re-thinking Model Inversion Attacks Against Deep Neural Networks
- CAD2RL: Real Single-Image Flight Without a Single Real Image
- Asymmetric Actor Critic for Image-Based Robot Learning
- Remotely Detectable Robot Policy Watermarking
- robosuite: A Modular Simulation Framework and Benchmark for Robot Learning
- MEDUSA: Medical Data Under Shadow Attacks via Hybrid Model Inversion
- Querying Kernel Methods Suffices for Reconstructing their Training Data
- Publishing Efficient On-device Models Increases Adversarial Vulnerability