PLUME: Building a Network-Native Foundation Model for Wireless Traces via Protocol-Aware Tokenization

Swadhin Pradhan; Shazal Irshad; Jerome Henry

cs.LGMar 13, 2026

PLUME: Building a Network-Native Foundation Model for Wireless Traces via Protocol-Aware Tokenization

Authors: Swadhin Pradhan, Shazal Irshad, Jerome Henry

Organizations: Cisco Systems, USA

Abstract

Foundation models succeed when they learn in the native structure of a modality, whether morphology-respecting tokens in language or pixels in vision. Wireless packet traces deserve the same treatment: meaning emerges from layered headers, typed fields, timing gaps, and cross-packet state machines, not flat strings. We present Plume (Protocol Language Understanding Model for Exchanges), a compact 140M-parameter foundation model for 802.11 traces that learns from structured PDML dissections. A protocol-aware tokenizer splits along the dissector field tree, emits gap tokens for timing, and normalizes identifiers, yielding 6.2x shorter sequences than BPE with higher per token information density. Trained on a curated corpus, Plume achieves 74-97% next-packet token accuracy across five real-world failure categories and AUROC >= 0.99 for zero-shot anomaly detection. On the same prediction task, frontier LLMs (Claude Opus 4.6, GPT-5.4) score comparably despite receiving identical protocol context, yet Plume does so with > 600x fewer parameters, fitting on a single GPU at effectively zero marginal cost vs. cloud API pricing, enabling on-prem, privacy-preserving root cause analysis.

Figures & tables

Panel of Figure 1 — Figure 1. HDBSCAN-based curation. (a) Long-tail cluster sizes; beacon-dominated clusters contain thousands of near-identical frames. (b) Frames group by protocol function, validating clustering-based deduplication.

	netFound ( Guthula et al., 2023 )	Lens ( Li et al., 2024 )	NetGPT ( Meng et al., 2023 )	LLMcap ( Tulczyjew et al., 2024 )	Plume
Architecture	Enc.	Enc-Dec	Dec.	Enc.	Dec.
Parameters	640M	–	–	${\sim}$ 66M	140M
Objective	MLM	Span	CLM	MLM	CLM
Tokenization	Field	BPE	BPE	WordPiece	Field
Layer hierarchy	Partial	✗	✗	✗	✓
802.11 mgmt/ctrl	✗	✗	✗	✗	✓

Table 1. Design-space comparison of foundation models trained on networking data. MLM = Masked Language Modeling, CLM = Causal Language Modeling, Span = masked span prediction. ✓ = supported, ✗ = not supported, – = not reported.

Model	Token Acc.	Gen?	Anomaly?	Classify?
Plume (GPT)	0.831	✓	✓	✓
3-gram	0.257	✓	–	–
Most-Frequent	0.226	–	–	–
Random	0.000	✓	–	–
BERT (encoder)	–	–	–	✓

Table 2. Architecture and baseline comparison.

Tokenizer	Vocab	Avg Tok/Pkt	Entropy	Ratio
Plume (field-value)	69,842	326.8	7.61	16.2 $\times$
BPE on PDML	100,277	2,014.0	6.70	2.6 $\times$
Byte-level	256	5,294.7	4.75	1.0 $\times$
Flat (no layers)	1,959	332.9	7.55	15.9 $\times$
NetGPT-style	100,277	2,013.0	6.70	2.6 $\times$

Table 3. Tokenization ablation: Plume ’s protocol-aware tokenizer vs. alternatives. Compression ratio is relative to byte-level (higher is better).

Figure 2. Average tokens per packet. Plume ’s protocol-aware tokenizer yields

6.2\times

shorter sequences than BPE and

16.2\times

shorter than byte-level.

Metric	Train	Val
Files	7,890	2,023
Packets	149,238	38,669
Total tokens	48.9M	12.7M
Unique tokens	58,214	22,198
Avg packet length (tokens)	326.8	326.8
Token entropy (bits)	7.601	7.610

Table 4. Dataset statistics after HDBSCAN+MMR curation. Beacon fraction drops from

>

50% in raw captures to 4.7%, while token entropy remains high (7.6 bits).

Category	Files	Packets	Avg Len	Entropy
Bad Password	9,960	116,469	421.8	7.52
EAPOL Timeout	9,992	257,057	318.9	7.37
Invalid PMKID	2,252	43,224	387.5	7.67
Unable Handle STA	9,997	143,746	377.8	7.53
Rejected Temp.	9,987	187,871	358.3	7.58

Table 5. Per-test-category corpus statistics showing the diversity of the evaluation data.

Model	Params	Width	Token Acc.	Field Acc.
Small (12H/768D)	140M	768	0.903	0.884
Medium (16H/1024D)	225M	1024	0.927	0.922
Large (24H/1536D)	450M	1536	0.801	0.780

Table 6. Multi-model scaling: same 69K vocabulary and 12-layer depth; differences reflect model width.

Figure 3. Token accuracy by category for Small (140M), Medium (225M), and Large (450M). Same vocabulary and depth; differences reflect model width.

	Small	Medium	Large	GPT-5.2
Parameters	140M	225M	450M	N/A
Size (FP16)	280 MB	449 MB	901 MB	N/A
Peak VRAM	594 MB	937 MB	1,865 MB	N/A
Latency (mean)	280 ms	384 ms	666 ms	$\sim$ 2 s
Throughput	200.2 pkt/s	150.6 pkt/s	82.7 pkt/s	N/A
Cost/1K pkt	$\sim$ $0	$\sim$ $0	$\sim$ $0	$4.92

Table 7. Efficiency (single A10G, 24 GB) vs. GPT-5.2 API ($4.92/1K pkt at published pricing ( OpenAI, 2025a ) ). Peak VRAM includes PyTorch ( Paszke et al., 2019 ) overhead.

Figure 4. Per-field accuracy by category (left) and 10 best/worst fields (right). Addresses and frame control are near-perfect; timing and rare fields are hardest.

Figure 5. Prediction accuracy vs. context length. Accuracy saturates by 2–3 packets, matching typical 802.11 exchange length.

Figure 6. Cross-category generalization. (a) Per-layer accuracy: WLAN $>$ 97.5% across all categories. (b) Cosine similarity of accuracy profiles: categories with shared protocol flows cluster together.

Failure Category	Plume	Claude 4.6	GPT-5.4
Bad Password	0.960	0.854	0.790
EAPOL Timeout	0.915	0.904	0.863
Invalid PMKID	0.753	0.865	0.852
Unable to Handle New STA	0.868	0.903	0.862
Rejected Temporarily	0.958	0.938	0.920
Overall mean	0.891	0.893	0.857

Table 8. Next-packet token accuracy: Plume (Small, 140M) vs. frontier LLMs. Total of 682 prediction pairs across five categories. Best per category in bold .

Figure 7. Per-category token accuracy for Plume (Small, 140M) vs. Claude Opus 4.6 and GPT-5.4. Plume matches or exceeds frontier LLMs on stereotyped protocol flows while using

>

600

\times

fewer parameters.

Failure Category	Token Acc.	Field Acc.	PPL	$n$
Bad Password	0.973 $\pm$ 0.110	0.963	2.1	50
EAPOL Timeout	0.937 $\pm$ 0.182	0.959	2.3	50
Invalid PMKID	0.741 $\pm$ 0.299	0.957	2.3	50
Unable to Handle. STA	0.880 $\pm$ 0.111	0.963	2.1	50
Rejected Temporarily	0.969 $\pm$ 0.063	0.962	2.1	50

Table 9. Per-category prediction quality (50 PCAPs each;

\pm

denotes std across PCAPs).

Figure 8. Per-PCAP token accuracy distribution (

n{=}50

). Boxes: IQR; whiskers:

1.5\times

IQR; circles: outliers.

Figure 9. Per-protocol-layer prediction accuracy (mean

\pm

std across categories). EAPOL fields are predicted perfectly; WLAN exceeds 96%; OTHER reaches 93–95%.

Failure Category	Mean	CI Lower	CI Upper
Bad Password	0.973	0.942	0.996
EAPOL Timeout	0.937	0.886	0.977
Invalid PMKID	0.741	0.658	0.822
Unable to Handle New STA	0.880	0.849	0.912
Rejected Temporarily	0.969	0.952	0.986

Table 10. 95% bootstrap confidence intervals for per-category token accuracy (

n{=}50

PCAPs each).

Failure Category	Plume		Pkt-Length
	AUROC	AUPRC	AUROC	AUPRC
Bad Password	0.99	0.99	0.95	0.95
EAPOL Timeout	1.00	1.00	0.78	0.76
Invalid PMKID	1.00	1.00	0.68	0.75
Unable to Handle. STA	1.00	1.00	0.79	0.80
Rejected Temporarily	1.00	1.00	0.79	0.73

Table 11. Zero-shot anomaly detection (AUROC/AUPRC,

n{=}50

per category vs. healthy baselines).

Figure 10. Zero-shot anomaly detection. (a) ROC curves: AUROC

\geq

0.99 for all categories. (b) Mean per-token probability vs. healthy baseline (dashed).

Figure 11. Per-token probability: healthy (green) vs. failure (red). Failure captures consistently score lower.

Classifier	Accuracy	Macro F1	Best Category
Logistic Regression	0.68	0.68	Rejected Temp. (0.83)
Random Forest	0.73	0.73	Rejected Temp. (0.85)
kNN	0.62	0.61	Rejected Temp. (0.78)

Table 12. Root cause classification: three classifiers on 19-D per-PCAP features (

n{=}250

, 5 categories

\times

50).

Figure 12. Logistic Regression confusion matrix ( $n{=}250$ ). Most reliable: Rejected Temporarily, Unable to Handle New STA. Most confused: Invalid PMKID.

Figure 13. t-SNE of 19-D per-PCAP features colored by failure category. EAPOL-related categories partially overlap, reflecting shared protocol flows.

Wireless Foundation Models

Wireless foundation models learn reusable representations of radio channels, signals, network traces, and telemetry for transfer across wireless tasks and settings.

See all 32 papers

Protocol-Aware Tokenization and Architecture Co-Design for Wireless Packet Foundation Models
May 14, 2026Swadhin Pradhan, Shazal Irshad, Jerome HenryWireless Foundation Models Wireless Communications
Cisco Systems
Evidence-Grounded Ensemble Diagnosis of 802.11 Packet Captures: A Multi-Stage Pipeline with Deterministic Reliability Scoring
Jun 5, 2026Jerome Henry, Swadhin Pradhan, Miroslav PopovicLLM Evaluation Evidence-Grounded Reasoning
Cisco Systems
A Protocol-Language Model for Network Intrusion (Without Deep Packet Inspection)
May 29, 2026Vivek Kumar SharmaNetwork Intrusion Detection Encrypted Network Traffic Analysis
Palo Alto Networks

May 14, 2026J/K move · Enter open · S save

Protocol-Aware Tokenization and Architecture Co-Design for Wireless Packet Foundation Models

Swadhin Pradhan, Shazal Irshad, Jerome Henry

Cisco Systems

Wireless Foundation Models Wireless Communications Wireless Local Area Networks

Open paper AskPDF arXiv

What matters more for building foundation models for wireless packet traces: the tokenizer or the architecture or both? To answer this question, we build on PLUME Anonymous [2026], which introduced protocol-aware tokenization for 802.11 traces; we scale model depth and transfer the same tokenizer to a fundamentally different architecture family. A deeper GPT (PLUME-DEEP, 24 layers) reaches 98.2% top-1 accuracy, gaining 32 points over the original 12-layer design, while a Mamba-2 state-space variant (PLUME-MAMBA) achieves 96.1% with 1.7x higher throughput and 2x longer context. The key insight emerges from a controlled 2x2 comparison across tokenizers and architectures: changing the tokenizer swings accuracy by 32 points; changing the architecture moves it by only 2. Protocol-aware tokenization is the primary performance lever, and the backbone becomes a deployment knob trading accuracy for speed.

Abstract

Wireless Foundation Models

Explore similar work

Protocol-Aware Tokenization and Architecture Co-Design for Wireless Packet Foundation Models

Evidence-Grounded Ensemble Diagnosis of 802.11 Packet Captures: A Multi-Stage Pipeline with Deterministic Reliability Scoring

A Protocol-Language Model for Network Intrusion (Without Deep Packet Inspection)