{"id":3227,"date":"2025-09-11T14:01:28","date_gmt":"2025-09-11T14:01:28","guid":{"rendered":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/?p=3227"},"modified":"2025-09-11T14:31:59","modified_gmt":"2025-09-11T14:31:59","slug":"lightweight-federated-lora-sb-modulation-classifier-with-vision-llm","status":"publish","type":"post","link":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/?p=3227","title":{"rendered":"Lightweight Federated LoRA-SB Modulation Classifier with Vision-LLM"},"content":{"rendered":"\n<figure class=\"wp-block-embed is-type-wp-embed is-provider-spectrcyde wp-block-embed-spectrcyde\"><div class=\"wp-block-embed__wrapper\">\n<blockquote class=\"wp-embedded-content\" data-secret=\"Oiu3mt34cf\"><a href=\"https:\/\/172-234-197-23.ip.linodeusercontent.com\/?page_id=3224\">Lightweight Federated LoRA-SB SignalClassification with Vision-LLM<\/a><\/blockquote><iframe class=\"wp-embedded-content\" sandbox=\"allow-scripts\" security=\"restricted\" style=\"position: absolute; visibility: hidden;\" title=\"&#8220;Lightweight Federated LoRA-SB SignalClassification with Vision-LLM&#8221; &#8212; Spectrcyde\" src=\"https:\/\/172-234-197-23.ip.linodeusercontent.com\/?page_id=3224&#038;embed=true#?secret=Vcf9Q7KHLx#?secret=Oiu3mt34cf\" data-secret=\"Oiu3mt34cf\" width=\"600\" height=\"338\" frameborder=\"0\" marginwidth=\"0\" marginheight=\"0\" scrolling=\"no\"><\/iframe>\n<\/div><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\"><\/h3>\n\n\n\n<p>A Reproducible Simulation Report<br><strong>Benjamin J. Gilbert<\/strong><br>Spectrcyde RF Quantum SCYTHE, College of the Mainland<\/p>\n\n\n\n<p><strong>ORCID<\/strong>: <a href=\"https:\/\/orcid.org\/0009-0006-2298-6538\">0009-0006-2298-6538<\/a><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>We present a lightweight modulation classifier designed for edge-side federation. The scheme applies <strong>LoRA-SB rank updates<\/strong> in a federated setting (Fed-SB) and may optionally incorporate a <strong>vision-LLM channel<\/strong> that interprets spectrograms as image-like features.<\/p>\n\n\n\n<p>A <strong>closed benchmark harness<\/strong> is built: spectra are synthesized, figures and tables auto-generated, and real API endpoints stubbed (gRPC\/LLM) for safety and reproducibility. Even on a <strong>small CPU run<\/strong>, the system shows strong macro-F1 and clear per-class precision\u2013recall curves.<\/p>\n\n\n\n<p>This report emphasizes <strong>engineering reproducibility<\/strong> rather than field-grade accuracy or privacy proofs. Differentially-private SGD (DP-SGD) is included in the framework but disabled for the paper experiments.<\/p>\n\n\n\n<p><strong>Index Terms \u2014<\/strong> modulation classification; federated learning; parameter-efficient fine-tuning; LoRA; privacy; calibration<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h4 class=\"wp-block-heading\">Contribution and Implementation<\/h4>\n\n\n\n<p>We explore a <strong>deployable path for RF classifiers across heterogeneous edge devices<\/strong>.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Parameter-efficient LoRA-SB rank updates aggregated with <strong>FedAvg<\/strong>.<\/li>\n\n\n\n<li>Optional <strong>vision-LLM parsing<\/strong> to strengthen robustness from spectra.<\/li>\n\n\n\n<li>Evaluation includes <strong>temperature scaling<\/strong> for calibration and reliability.<\/li>\n\n\n\n<li>Implementation ships with: LoRA-SB layers, DP-SGD switch, gRPC aggregation stub, vision-LLM hook.<\/li>\n<\/ul>\n\n\n\n<p>Code is fully reproducible: one command trains, evaluates, and renders results. Framework is implemented in <strong>PyTorch<\/strong> with <strong>scikit-learn utilities<\/strong>.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><a href=\"https:\/\/mastodon.social\/@Bgilbert1984\/115186016452665344\"><img data-opt-id=321536872  fetchpriority=\"high\" decoding=\"async\" width=\"766\" height=\"391\" src=\"https:\/\/ml6vmqguit1n.i.optimole.com\/w:auto\/h:auto\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/09\/image-34.png\" alt=\"\" class=\"wp-image-3228\" srcset=\"https:\/\/ml6vmqguit1n.i.optimole.com\/w:766\/h:391\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/09\/image-34.png 766w, https:\/\/ml6vmqguit1n.i.optimole.com\/w:300\/h:153\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/09\/image-34.png 300w\" sizes=\"(max-width: 766px) 100vw, 766px\" \/><\/a><\/figure>\n<\/div>\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-large\"><a href=\"https:\/\/www.facebook.com\/share\/p\/1P6jPEtSCk\/\"><img data-opt-id=421991779  fetchpriority=\"high\" decoding=\"async\" width=\"683\" height=\"1024\" src=\"https:\/\/ml6vmqguit1n.i.optimole.com\/w:683\/h:1024\/q:mauto\/f:best\/http:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/08\/ChatGPT-Image-Aug-30-2025-07_05_10-PM.png\" alt=\"\" class=\"wp-image-3048\" srcset=\"https:\/\/ml6vmqguit1n.i.optimole.com\/w:683\/h:1024\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/08\/ChatGPT-Image-Aug-30-2025-07_05_10-PM.png 683w, https:\/\/ml6vmqguit1n.i.optimole.com\/w:200\/h:300\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/08\/ChatGPT-Image-Aug-30-2025-07_05_10-PM.png 200w, https:\/\/ml6vmqguit1n.i.optimole.com\/w:768\/h:1152\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/08\/ChatGPT-Image-Aug-30-2025-07_05_10-PM.png 768w, https:\/\/ml6vmqguit1n.i.optimole.com\/w:1024\/h:1536\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/08\/ChatGPT-Image-Aug-30-2025-07_05_10-PM.png 1024w\" sizes=\"(max-width: 683px) 100vw, 683px\" \/><\/a><\/figure>\n<\/div>\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"has-text-align-center\">\u2699\ufe0f <strong>Guangdong framing<\/strong> \u2192 pragmatic, reproducibility-first, small-form factor ready, echoing the &#8220;Shenzhen-style&#8221; engineering culture: fast iteration, edge-deploy focus, no wasted words, and a clear pipeline from <strong>bench \u2192 shop floor \u2192 deployment<\/strong>.<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><a href=\"https:\/\/mastodon.social\/@Bgilbert1984\/115170225230187883\"><img data-opt-id=112362826  data-opt-src=\"https:\/\/ml6vmqguit1n.i.optimole.com\/w:auto\/h:auto\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/09\/image-35.png\"  decoding=\"async\" width=\"421\" height=\"404\" src=\"data:image/svg+xml,%3Csvg%20viewBox%3D%220%200%20100%%20100%%22%20width%3D%22100%%22%20height%3D%22100%%22%20xmlns%3D%22http%3A%2F%2Fwww.w3.org%2F2000%2Fsvg%22%3E%3Crect%20width%3D%22100%%22%20height%3D%22100%%22%20fill%3D%22transparent%22%2F%3E%3C%2Fsvg%3E\" alt=\"\" class=\"wp-image-3229\" old-srcset=\"https:\/\/ml6vmqguit1n.i.optimole.com\/w:421\/h:404\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/09\/image-35.png 421w, https:\/\/ml6vmqguit1n.i.optimole.com\/w:300\/h:288\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/09\/image-35.png 300w\" \/><\/a><\/figure>\n<\/div>\n\n\n<h3 class=\"wp-block-heading\">III. DATASET AND FEATURES<\/h3>\n\n\n\n<p>We construct a controlled synthetic dataset of <strong>band-limited spectra<\/strong> across eight representative classes: <strong>AM, FM, SSB, CW, PSK, FSK, NOISE, and UNKNOWN<\/strong>. Each instance is generated with parameterized modulation settings, ensuring diversity in bandwidth, carrier offset, and noise floor.<\/p>\n\n\n\n<p>For each synthetic spectrum, we compute a compact feature vector: <strong>analytic descriptors<\/strong> such as estimated bandwidth, spectral flatness, and roll-off slope. In addition, we enable an <strong>optional visual side channel<\/strong>, 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 <strong>identical runs on any machine without external dependencies<\/strong>.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">IV. TRAINING AND EVALUATION<\/h3>\n\n\n\n<p>Training follows a <strong>single local epoch per client<\/strong> under the <strong>Fed-SB framework<\/strong>. Each client updates only the <strong>low-rank R matrices<\/strong> of LoRA-SB layers, while base parameters (A and B) remain frozen. Aggregation is implemented via <strong>weighted averaging across clients<\/strong>, consistent with FedAvg.<\/p>\n\n\n\n<p>Evaluation is performed on a <strong>held-out synthetic split<\/strong>. The pipeline reports:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Accuracy and macro-F1<\/strong> as global indicators of balance.<\/li>\n\n\n\n<li><strong>Per-class precision, recall, and F1<\/strong> (Table II).<\/li>\n\n\n\n<li><strong>Confusion matrices and precision\u2013recall curves<\/strong>, automatically exported for inspection.<\/li>\n<\/ul>\n\n\n\n<p>All artifacts are reproducibly emitted into the <strong>figures\/<\/strong>, <strong>metrics\/<\/strong>, and <strong>tex\/<\/strong> directories. Tables and figures can be regenerated with a single command, ensuring alignment with publication requirements.<\/p>\n\n\n\n<p><strong>TABLE I<\/strong> summarizes the overall classifier statistics. Despite the small CPU-bound run, Fed-SB achieves stable accuracy and interpretable macro-F1. <strong>TABLE II<\/strong> reports detailed per-class performance. Predictably, some classes (PSK, NOISE) achieve non-trivial discrimination, while others (SSB, FSK) collapse under the current synthetic configuration.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">V. RESULTS<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">A. Classification Performance<\/h4>\n\n\n\n<p>The classifier demonstrates <strong>moderate success in separating structured digital modulations<\/strong> (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.<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">B. Federated Learning<\/h4>\n\n\n\n<p>Federated training with <strong>LoRA-SB rank updates<\/strong> 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 <strong>bandwidth-limited edge deployment scenarios<\/strong> typical in IoT and tactical RF networks.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">VI. IMPLEMENTATION NOTES<\/h3>\n\n\n\n<p>The prototype module defines a <code>LoRASBLayer<\/code> inserted into a compact MLP backbone. Each simulated client:<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>Trains its local R matrices for one epoch.<\/li>\n\n\n\n<li>Packages updates into a gRPC message (stubbed during paper runs).<\/li>\n\n\n\n<li>Aggregator performs weighted averaging of R updates.<\/li>\n<\/ol>\n\n\n\n<p>An optional <strong>vision-LLM endpoint<\/strong> 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.<\/p>\n\n\n\n<p>The codebase includes a <strong>DP-SGD implementation<\/strong> 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.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">VII. SCOPE AND LIMITATIONS<\/h3>\n\n\n\n<p>This study is <strong>synthetic-only<\/strong>. Reported results should be interpreted as a <strong>reproducibility baseline<\/strong>, not as field-validated performance. No claim is made regarding on-air generalization, robustness under real SNR\/channel dynamics, or strict privacy guarantees.<\/p>\n\n\n\n<p>Both the gRPC and vision-LLM integrations are exercised only through <strong>safe stubs<\/strong>, ensuring the build remains self-contained. This design choice reflects a <strong>Guangdong-style engineering ethos<\/strong>: prototype fast, guarantee reproducibility, keep the system minimal and portable.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">VIII. FUTURE WORK<\/h3>\n\n\n\n<p>Key extensions include:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Integration of <strong>formal DP accounting<\/strong> for federated privacy claims.<\/li>\n\n\n\n<li>Running <strong>real federated rounds across devices<\/strong>, with heterogeneous data splits.<\/li>\n\n\n\n<li>Conducting <strong>channel-aware sweeps<\/strong> across SNR and multipath profiles.<\/li>\n\n\n\n<li>Expanding the spectrogram parser from stubbed JSON to a <strong>live vision-LLM pipeline<\/strong>, enabling adaptive feature extraction in noisy or distorted bands.<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>\u2699\ufe0f <strong>Summary:<\/strong> 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 <strong>deployability across constrained devices<\/strong>, a hallmark of pragmatic edge AI design.<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><a href=\"https:\/\/mastodon.social\/@Bgilbert1984\/115159942769065202\"><img data-opt-id=504650587  data-opt-src=\"https:\/\/ml6vmqguit1n.i.optimole.com\/w:auto\/h:auto\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/09\/image-36.png\"  decoding=\"async\" width=\"724\" height=\"184\" src=\"data:image/svg+xml,%3Csvg%20viewBox%3D%220%200%20100%%20100%%22%20width%3D%22100%%22%20height%3D%22100%%22%20xmlns%3D%22http%3A%2F%2Fwww.w3.org%2F2000%2Fsvg%22%3E%3Crect%20width%3D%22100%%22%20height%3D%22100%%22%20fill%3D%22transparent%22%2F%3E%3C%2Fsvg%3E\" alt=\"\" class=\"wp-image-3233\" old-srcset=\"https:\/\/ml6vmqguit1n.i.optimole.com\/w:724\/h:184\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/09\/image-36.png 724w, https:\/\/ml6vmqguit1n.i.optimole.com\/w:300\/h:76\/q:mauto\/f:best\/https:\/\/172-234-197-23.ip.linodeusercontent.com\/wp-content\/uploads\/2025\/09\/image-36.png 300w\" \/><\/a><\/figure>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>A Reproducible Simulation ReportBenjamin J. GilbertSpectrcyde RF Quantum SCYTHE, College of the Mainland ORCID: 0009-0006-2298-6538 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:&hellip;&nbsp;<a href=\"https:\/\/172-234-197-23.ip.linodeusercontent.com\/?p=3227\" rel=\"bookmark\"><span class=\"screen-reader-text\">Lightweight Federated LoRA-SB Modulation Classifier with Vision-LLM<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":2467,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"neve_meta_sidebar":"","neve_meta_container":"","neve_meta_enable_content_width":"","neve_meta_content_width":0,"neve_meta_title_alignment":"","neve_meta_author_avatar":"","neve_post_elements_order":"","neve_meta_disable_header":"","neve_meta_disable_footer":"","neve_meta_disable_title":"","footnotes":""},"categories":[10],"tags":[],"class_list":["post-3227","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-signal_scythe"],"_links":{"self":[{"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=\/wp\/v2\/posts\/3227","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=3227"}],"version-history":[{"count":4,"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=\/wp\/v2\/posts\/3227\/revisions"}],"predecessor-version":[{"id":3236,"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=\/wp\/v2\/posts\/3227\/revisions\/3236"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=\/wp\/v2\/media\/2467"}],"wp:attachment":[{"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=3227"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=3227"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/172-234-197-23.ip.linodeusercontent.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=3227"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}