Description
Adversarial training pits a primary model against an adversary that probes its weaknesses, then optimises the primary against the worst-case adversary. In WiFi sensing two flavours dominate: (a) domain-adaptation via a domain-discriminator adversary (DANN-style) and (b) robustness training against perturbations of CSI inputs. It is conceptually distinct from generative-adversarial-network-style sample synthesis.
When it's used
- Domain-invariant feature learning for cross-environment CSI
- Robustness against CSI perturbations / hardware noise
- Defending fingerprinting against replay attacks
Limitations
- Optimisation is unstable; collapse modes are common
- Adversary design strongly affects what gets learned
- Robustness gains often trade against clean accuracy