Lightweight Federated LoRA-SB Modulation Classifier with Vision-LLM

Lightweight Federated LoRA-SB SignalClassification with Vision-LLM

A Reproducible Simulation Report
Benjamin J. Gilbert
Spectrcyde RF Quantum SCYTHE, College of the Mainland

We present a lightweight modulation classifier designed for edge-side federation. The scheme applies LoRA-SB rank updates in a federated setting (Fed-SB) and may optionally incorporate a vision-LLM channel that interprets spectrograms as image-like features.

A closed benchmark harness is built: spectra are synthesized, figures and tables auto-generated, and real API endpoints stubbed (gRPC/LLM) for safety and reproducibility. Even on a small CPU run, the system shows strong macro-F1 and clear per-class precision–recall curves.

This report emphasizes engineering reproducibility rather than field-grade accuracy or privacy proofs. Differentially-private SGD (DP-SGD) is included in the framework but disabled for the paper experiments.

Index Terms — modulation classification; federated learning; parameter-efficient fine-tuning; LoRA; privacy; calibration

Contribution and Implementation

We explore a deployable path for RF classifiers across heterogeneous edge devices.

Parameter-efficient LoRA-SB rank updates aggregated with FedAvg.
Optional vision-LLM parsing to strengthen robustness from spectra.
Evaluation includes temperature scaling for calibration and reliability.
Implementation ships with: LoRA-SB layers, DP-SGD switch, gRPC aggregation stub, vision-LLM hook.

Code is fully reproducible: one command trains, evaluates, and renders results. Framework is implemented in PyTorch with scikit-learn utilities.

⚙️ Guangdong framing → pragmatic, reproducibility-first, small-form factor ready, echoing the “Shenzhen-style” engineering culture: fast iteration, edge-deploy focus, no wasted words, and a clear pipeline from bench → shop floor → deployment.

III. DATASET AND FEATURES

We construct a controlled synthetic dataset of band-limited spectra across eight representative classes: AM, FM, SSB, CW, PSK, FSK, NOISE, and UNKNOWN. Each instance is generated with parameterized modulation settings, ensuring diversity in bandwidth, carrier offset, and noise floor.

For each synthetic spectrum, we compute a compact feature vector: analytic descriptors such as estimated bandwidth, spectral flatness, and roll-off slope. In addition, we enable an optional visual side channel, where a spectrogram is parsed into structured features. To preserve reproducibility, the spectrogram parser is stubbed; in the simulation it consistently returns JSON objects containing fields such as peak count, centroid symmetry, and band energy distribution. This guarantees identical runs on any machine without external dependencies.

IV. TRAINING AND EVALUATION

Training follows a single local epoch per client under the Fed-SB framework. Each client updates only the low-rank R matrices of LoRA-SB layers, while base parameters (A and B) remain frozen. Aggregation is implemented via weighted averaging across clients, consistent with FedAvg.

Evaluation is performed on a held-out synthetic split. The pipeline reports:

Accuracy and macro-F1 as global indicators of balance.
Per-class precision, recall, and F1 (Table II).
Confusion matrices and precision–recall curves, automatically exported for inspection.

All artifacts are reproducibly emitted into the figures/, metrics/, and tex/ directories. Tables and figures can be regenerated with a single command, ensuring alignment with publication requirements.

TABLE I summarizes the overall classifier statistics. Despite the small CPU-bound run, Fed-SB achieves stable accuracy and interpretable macro-F1. TABLE II reports detailed per-class performance. Predictably, some classes (PSK, NOISE) achieve non-trivial discrimination, while others (SSB, FSK) collapse under the current synthetic configuration.

V. RESULTS

A. Classification Performance

The classifier demonstrates moderate success in separating structured digital modulations (PSK, NOISE) but struggles with closely-spaced analog modes (SSB, FSK). This mirrors real-world deployment constraints: synthetic-only training cannot capture the richness of fading channels, oscillator instabilities, or multipath distortions.

B. Federated Learning

Federated training with LoRA-SB rank updates converges efficiently. Figure 4 shows accuracy plotted against communication cost across federated rounds. By transmitting only the R matrices, communication is minimized while maintaining convergence properties. The framework thus aligns with bandwidth-limited edge deployment scenarios typical in IoT and tactical RF networks.

VI. IMPLEMENTATION NOTES

The prototype module defines a LoRASBLayer inserted into a compact MLP backbone. Each simulated client:

Trains its local R matrices for one epoch.
Packages updates into a gRPC message (stubbed during paper runs).
Aggregator performs weighted averaging of R updates.

An optional vision-LLM endpoint is exposed, returning JSON with features such as bandwidth, peak layout, and symmetry descriptors. For reproducibility, this endpoint is stubbed with fixed JSON templates, eliminating dependency on external APIs.

The codebase includes a DP-SGD implementation via Opacus. Although present, it is disabled by default in paper runs to maintain runtime stability and reproducibility. Switching privacy on or off is a single configuration toggle.

VII. SCOPE AND LIMITATIONS

This study is synthetic-only. Reported results should be interpreted as a reproducibility baseline, not as field-validated performance. No claim is made regarding on-air generalization, robustness under real SNR/channel dynamics, or strict privacy guarantees.

Both the gRPC and vision-LLM integrations are exercised only through safe stubs, ensuring the build remains self-contained. This design choice reflects a Guangdong-style engineering ethos: prototype fast, guarantee reproducibility, keep the system minimal and portable.

VIII. FUTURE WORK

Key extensions include:

Integration of formal DP accounting for federated privacy claims.
Running real federated rounds across devices, with heterogeneous data splits.
Conducting channel-aware sweeps across SNR and multipath profiles.
Expanding the spectrogram parser from stubbed JSON to a live vision-LLM pipeline, enabling adaptive feature extraction in noisy or distorted bands.

⚙️ Summary: This paper provides a reproducible, minimal simulation harness for federated modulation classification using LoRA-SB rank updates. While limited to synthetic benchmarks, the architecture demonstrates deployability across constrained devices, a hallmark of pragmatic edge AI design.