Description
The umbrella concern that a CSI / wireless-sensing model performs well not just on the test split of the data on which it was trained but on perturbed, shifted, or adversarial inputs encountered in deployment. Robustness covers benign perturbations (noise, missing subcarriers, antenna failures); generalization covers structural shifts (new users, new rooms, new hardware). Both are operationalized through evaluation protocols (leave-one-environment-out, cross-user splits) rather than by any single model property, and both feed directly into the environment-dependence research concern.
Why it's hard
- Standard validation splits (random per-sample) wildly overstate deployment performance.
- Defining "the relevant distribution shift" is itself a modeling decision; benchmarks differ.
- Robustness against one perturbation often trades off against robustness against another.
- Reproducibility crises in wireless sensing complicate cross-paper comparisons.
- Theoretical robustness/generalization bounds for deep CSI models are far from tight.
Common approaches
- Cross-environment, cross-user, cross-device evaluation protocols.
- Adversarial training and input augmentation.
- Test-time adaptation; entropy minimization at inference.
- Standardized benchmarks (SDP, SenseFi) to enable apples-to-apples comparison.
Source Papers
- wang2026_2758 ↗ — Wi-Fi sensing generalizability survey (the canonical taxonomy).
- guarino2026_e72c ↗ — CSI-based WiFi sensing datasets + reproducibility focus.
- zhang2026_ccac ↗ — SDP benchmarking framework for reproducible wireless sensing.
- yang2023_a34a ↗ — SenseFi library and benchmark for deep WiFi human sensing.
- koo2026_a08d ↗ — green wireless sensing survey (robustness vs efficiency trade-offs).