Feasibility of finer-grained tokenization for Transformer-based full-waveform LiDAR models under jamming

Determine whether adopting finer-grained tokenization in Transformer-based models that operate on LiDAR full-waveform data—specifically Neural DSP—can mitigate performance degradation under jamming attacks that produce excessive waveform peaks, while keeping computational and memory requirements within practical limits for real-time deployment.

Background

The paper evaluates PULSAR-Net against baselines including Neural DSP, a Transformer-based method that processes full-waveform LiDAR data using tokenization. Under jamming attacks, full waveforms contain many peaks, and the authors observe that Neural DSP’s current tokenization granularity is insufficient to handle this scenario.

They note that using more fine-grained tokenization might address the observed limitation but caution that finer granularity increases computational and memory costs. The authors explicitly state that whether such finer tokenization is viable remains an open question, highlighting a trade-off between representational fidelity and resource constraints.

References

This is because Neural DSP is not designed for jamming attack scenarios, where the full waveform contains an excessive number of peaks that exceed the capacity of its current tokenization granularity. Finer-grained tokenization may alleviate this limitation, but remains an open question given the associated increases in computational and memory requirements.

— Neural Reconstruction of LiDAR Point Clouds under Jamming Attacks via Full-Waveform Representation and Simultaneous Laser Sensing  (2604.00371 - Yoshida et al., 1 Apr 2026) in Section 5.2 (Results)