UIUC Fixes DeepONet's Catastrophic Fragility, Slashing Errors by 87%
A hybrid defense combining active learning with a learnable input bottleneck cuts adversarial error on Burgers equation surrogates by 87 percent.
Takeaways
- A combined active learning plus input denoising defense cuts neural operator adversarial error 87% on the Burgers equation.
- Standard DeepONet training degrades 3.7x under attack: 3.27% clean error to 12.15% under adversarial perturbation.
- Combined method achieves 1.21% baseline plus 0.83% robustness error under a fixed 600-simulation budget.
- Key twist: active learning uses physics-corrected labels from the solver, not adversarial labels, avoiding accuracy-robustness trade-off.
- Architecture adds a learnable 64 to 32 to 64 autoencoder bottleneck before DeepONet's branch network to filter perturbations.
- Authors hypothesize uniform data sampling is suboptimal because each architecture concentrates sensitivity in a different subspace.
Neural operators like DeepONet have become the go-to fast surrogate for physics simulations, replacing hour-long finite element runs with millisecond inference. A growing body of work shows they hide a nasty flaw: they can look perfect on validation data and still collapse when the input is nudged by a barely perceptible perturbation. A new paper from researchers at UIUC and IIT Delhi proposes a two-pronged defense that closes most of that gap on a canonical PDE benchmark.
The paper combines active learning that hunts for a model's weak spots with a small autoencoder bottleneck bolted onto the front of the network. Together they drop combined error from 15.42% to 2.04% on the viscous Burgers equation, an 87% reduction that neither technique achieves alone.
Why neural operators break under tiny nudges
A neural operator learns the mapping from an input function (say, an initial condition) to a solution function (the state at some later time). Prior work by the same group established that these models are catastrophically fragile: on a multi-physics CFD benchmark, models with relative L2 errors below 10 to the minus 5 on clean validation data suffered five orders of magnitude degradation under sparse perturbations affecting fewer than 1% of input dimensions.
The physics itself does not misbehave. When the same perturbed input is fed to both the neural operator and the physics solver, the solver produces a solution close to the unperturbed one, while the neural operator's prediction diverges catastrophically. The learned mapping amplifies noise that the underlying PDE would smooth out. Worse, 100% of successful single-point attacks evaded standard anomaly detection, meaning conventional monitoring pipelines would never see them coming. For digital twins running on nuclear reactors, that failure mode is disqualifying.