Jev-IDS: System One Models for Network Intrusion Detection

Paulo Severo; Silvio E. Quincozes; Amanda Dias

cs.CROct 1, 2026

Jev-IDS: System One Models for Network Intrusion Detection

Authors: Paulo Severo, Silvio E. Quincozes, Amanda Dias

Organizations: Graduate Program in Software Engineering, Federal University of Pampa (UNIPAMPA), Alegrete, Rio Grande do Sul, Brazil

Abstract

Machine-learning Network Intrusion Detection Systems (IDS) depend on substantial labeled datasets and task-specific training, whereas Large Language Models (LLMs) detection can analyze flow records directly but incurs higher inference cost and latency, with less constrained outputs. This paper presents JEV-IDS, an open experimental general NIDS based on the Jev System One Model (SOM) to detect zero day intrusions Under label scarcity. JEV-IDS serializes one flow per request and asks JEV two questions: a binary attack probability and a finite-choice traffic category. Our results show that, at k=1, JEV was 4.8 times faster and 3.8 times cheaper than GPT-5.6 Luna, with 1.5 times higher novel-attack recall; it also produced 15 times fewer false alarms than a low-data Random Forest. Across 5,400 decisions on a 300-flow NSL-KDD pilot split, JEV achieved F1-Score 0.859, precision 0.941, recall 0.790, and novel-attack recall 0.838. Increasing k to 2 reduced its F1-Score to 0.839.

Figures & tables

Figure 1: Architecture of Jev-IDS . A target Flow is represented by its feature values in Card order and inserted into a Jev request together with the request instructions, column names, Category descriptions, and optional labeled Examples. Jev answers an intrusion-probability question and a Category question, from which Jev-IDS derives the corresponding detection outputs.

Flows	Normal	DoS	Probe	R2L	U2R
2,000	874	642	213	248	23

Table 1: Composition of the NSL-KDD evaluation set.

Detector	$k$	F1	Precision	Recall	Known recall	Novel recall	PR-AUC	ROC-AUC
Jev	0	$0.782\pm 0.005$	0.972	0.654	0.665	0.622	0.940	0.925
	1	$0.856\pm 0.025$	0.953	0.778	0.790	0.747	0.952	0.950
	2	$0.846\pm 0.009$	0.949	0.763	0.772	0.738	0.949	0.944
	4	$0.856\pm 0.015$	0.944	0.784	0.803	0.731	0.959	0.959
	8	$0.854\pm 0.017$	0.942	0.783	0.805	0.721	0.964	0.962
Gemini 3.6 Flash	0	$0.867\pm 0.005$	0.946	0.801	0.842	0.687	0.949	0.959

Table 2: Predictive results on the 2,000-Flow NSL-KDD evaluation set. F1 is reported as mean

\pm

sample standard deviation across three Example seeds. The remaining metrics are seed means.

Subset	$k$	Pairs	Discordant	Jev correct	Gemini correct	Exact $p$
All Flows	0	6,000	833	213	620	$<0.001$
All Flows	1	6,000	508	195	313	$<0.001$
All Flows	2	6,000	515	199	316	$<0.001$
All Flows	4	6,000	506	178	328	$<0.001$
All Flows	8	6,000	472	159	313	$<0.001$
Known attacks	0	2,478	489	25	464	$<0.001$

Table 3: Paired comparison between Jev and Gemini 3.6 Flash. The “Jev correct” and “Gemini correct” columns report the correctly classified instances among discordant decisions.

$k$	Jev cost	Gemini cost	Cost ratio	Jev latency	Gemini latency	Latency ratio
0	US$43	US$1,068	$25\times$	310 ms	2.24 s	$7.2\times$
1	US$74	US$1,651	$22\times$	315 ms	2.42 s	$7.7\times$
2	US$106	US$2,409	$23\times$	308 ms	2.65 s	$8.6\times$
4	US$169	US$2,869	$17\times$	322 ms	2.13 s	$6.6\times$
8	US$295	US$3,035	$10\times$	335 ms	1.99 s	$5.9\times$

Table 4: Operational comparison between Jev and Gemini 3.6 Flash. Cost denotes the estimated list price per one million Flow decisions, and latency denotes the mean wall-clock duration of a successful model request. Ratios are Gemini/ Jev .

Explore similar work

A Protocol-Language Model for Network Intrusion (Without Deep Packet Inspection)
May 29, 2026Vivek Kumar SharmaIntrusion Detection Network Traffic
Palo Alto Networks
Traffic-Aware Randomized Smoothing for LLM-Based Network Intrusion Detection
Jul 15, 2026Zhenpeng LiIntrusion Detection Smoothing
GenTI: Benchmarking LLMs for Autonomous IDPS Rule Generation for Unseen Attacks
Jun 4, 2026Hassan Jalil Hadi, Rehana Yasmin, Ali ShokerIntrusion Detection Cyber Threat Intelligence
Cyber Security and Resilience Technology (CyberSaR), King Abdullah University of Science and Technology (KAUST), Saudi Arabia

May 29, 2026, cs.CRJ/K move · Enter open · S save

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

Vivek Kumar Sharma

Palo Alto Networks

Intrusion Detection Network Traffic Language Modeling Roc-Auc Pretraining Lidar

Open paper AskPDF arXiv

Modern network intrusion detection systems (NIDS) are caught in a structural contradiction: the protocols carrying the highest threat intelligence are precisely those encrypted under TLS 1.3 and QUIC, where payload inspection yields nothing. We ask a simpler question -- what if the attack signature is not in the bytes, but in the rhythm? -- and answer it by treating network flows as a language whose grammar is written entirely in L3/L4 packet metadata: length, inter-arrival time, TTL, TCP flags, and hashed port numbers. We present PLM-NIDS, which proves three claims in sequence. (1) The grammar exists and is learnable: a RWKV-4 state-space model trained on 344,232 unlabelled Monday flows achieves a causal LM validation loss of 0.204, demonstrating that benign traffic has predictable, statistically consistent structure. (2) Attacks violate this grammar: the per-flow perplexity score cleanly separates benign from attack flows with PR-AUC = 0.93 using zero attack labels at training time. (3) This separation is architecturally nontrivial: an LSTM trained on identical token sequences degenerates to a majority-class predictor (ROC-AUC approximately 0.50, F1 = 0.91 by always predicting "attack"), proving that RWKV's causal pre-training provides an inductive bias unavailable to direct classifiers. Supervised fine-tuning further raises PR-AUC to 0.94 and ROC-AUC to 0.75, with a precision of 97.7% at the calibrated operating threshold. The RWKV backbone's O(T) recurrent inference enables per-packet streaming without flow buffering, making PLM-NIDS operationally viable at line rate. Because it reads only IP/TCP/UDP headers, it is inherently encryption-agnostic: TLS 1.3, QUIC, and future encrypted protocols are handled transparently.

Abstract

Explore similar work

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

Traffic-Aware Randomized Smoothing for LLM-Based Network Intrusion Detection

GenTI: Benchmarking LLMs for Autonomous IDPS Rule Generation for Unseen Attacks