RF monitoring stacks must compress and interpret highrate spectra under tight latency and power budgets. We target
the common case where the front-end produces windowed
FFT power spectra (magnitude-only) and the back-end must
(1) compress to a compact token sequence for downstream
classifiers, and (2) preserve class-relevant detail. We explore
multi-head linear attention (MHLA) with FlashAttention-style
backends and token-dropout as a simple, hardware-friendly
compressor.
Modern RF environments present increasingly complex
challenges beyond basic compression and classification. Emerging threats include “ghost” anomalies—stealthy emissions,
frequency-hopping signals, and sophisticated spoofing attacks—that evade traditional detection methods. These challenges are exacerbated in tactical edge deployments where
resource constraints limit processing capabilities. Our work
addresses these challenges by enabling:
- Multi-modal intelligence fusion: Compressed RF representations that maintain coherence with other sensor
modalities (visual, acoustic) - Scalable band monitoring: Processing up to 40% more
concurrent frequency bands on the same hardware through
efficient compression - Distribution-aware learning: Adaptive positional encoding via dynamic-θ RoPE to address signal characteristic
variations across diverse bands (ISM, cellular, GNSS,
aero)
We contribute: (1) an analysis of compression–accuracy
trade-offs using token-dropout in RF spectrum encoding; (2)
latency profiles across attention backends with varying token
counts; (3) an ablation study on positional encoding schemes;
and (4) integration of anomaly detection capabilities without
significant latency increases. Figure 1 presents the overall
architecture of our SpectrumEncoder system and its integration
into a complete SIGINT pipeline.