E-SpecFormer slims RF monitoring to 10k parameters, 92μs inference

Monitoring radio frequency (RF) spectrum in real-time on resource-constrained devices has been a challenge due to the high computational demands of deep learning models. A team of researchers from Sorbonne University and LIP6 introduces E-SpecFormer (Edge Spectrum monitoring Transformer), a family of scalable transformers designed for end-to-end automatic modulation recognition and covert channel (CC) detection. The key innovation is LiTAN (Linear Tanh Attention Network), a Softmax- and LayerNorm-free attention mechanism that reduces computational complexity while paradoxically increasing accuracy on RF tasks. E-SpecFormer comes in four variants—Nano, Small, Medium, Large—to suit different hardware constraints, from microcontrollers to FPGAs.

Performance results are striking, especially for the Nano variant. On the RadioML2018 dataset, it achieves 86.5% average accuracy for signal-to-noise ratios above 0 dB. On a hardware Trojan-based covert channel dataset, accuracy reaches 94.2%. All this with fewer than 10,000 parameters and an inference speed of just 92 microseconds per frame on FPGA/CPU co-execution—far faster and cheaper than current state-of-the-art edge models. These results position E-SpecFormer as a practical solution for real-time spectrum intelligence on Internet of Things (IoT) devices, enabling applications like spectrum monitoring, interference detection, and security threat identification at the network edge.