Skip to content

Flash-Attention MHLA for RF SpectrumCompression: SpectrumEncoder with Token-Dropout and RoPEAblations

We present a lightweight SpectrumEncoder for
compressing FFT power spectra using multi-head linear attention
(MHLA) with FlashAttention backends and token-dropout. We
report compression–accuracy trade-offs, latency profiles, and an
ablation on Rotary Positional Embeddings (RoPE). The method is
designed for real-time SIGINT pipelines where millisecond-level
latency and energy budgets matter, enabling up to 40% more
concurrent RF bands on the same hardware.