Sigma Defense
Telling drones apart by their radio signals
- 73.6%classification accuracy
- +16%over baseline models
- 0.714F1 score, up 0.157
I built a model that separates RF signals by drone class, to detect foreign or dangerous drones in conflict zones. It fuses spectrogram and power spectral density views of more than a million noisy signals using SimCLR, reaching 73.6% accuracy, 16% above the baselines.
To make detection hold up in noisy, non-stationary RF environments, I added a per-pixel sigmoid gate that ignores raw signal strength. That raised F1 by 0.157, to 0.714.
I also set up a ClearML experiment tracking pipeline across 10,000+ RF samples, which made comparing models much faster.