Transferability of robustness findings across model backbones

Determine whether the robustness findings obtained from evaluating input dropout, adversarial training, and self-supervised pretraining on the TST backbone transfer across alternative virtual-sensing model backbones.

Background

The study evaluates three robustification strategies—input sensor dropout (ISensD), projected-gradient-descent adversarial training (PGD), and failure-focused self-supervised pretraining (F2F)—using the TST architecture as a shared backbone. Although these methods improve corruption robustness in several settings, they also incur nominal-performance costs, and the paper does not establish whether the observed trade-offs and relative effectiveness persist for other architectures such as gradient-boosted trees, convolutional models, recurrent models, or MLP-mixing models.

The unresolved issue is therefore the external validity of the reported robustness conclusions with respect to backbone architecture. Establishing transferability would clarify whether the results reflect generally effective robustification principles or interactions specific to TST.

References

Second, we evaluate three representative robustification methods on a shared TST backbone to enable a fair head-to-head comparison with F2F; broader families of defenses, including techniques from the vision-robustness literature, and the question of whether our findings transfer across backbones remain open, and we plan to address both in future work using MuViS-C as the shared testbed.

— Reliable Virtual Sensing: A Multi-Domain Benchmark for Robustness Under Sensor Failures  (2609.18396 - Brandt et al., 16 Sep 2026) in Section 5, “Limitations and future work”