<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/looking-for-lateral-movement-with-a-neural-network-trained-on-synthetic-data-f7vvndwzi" -->

---
title: Looking for lateral movement with a neural network...
description: A detailed technical experiment trains neural networks to detect lateral movement in corporate networks using entirely synthetic training data generated from a...
canonical: https://daily.dev/posts/looking-for-lateral-movement-with-a-neural-network-trained-on-synthetic-data-f7vvndwzi
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: Looking for lateral movement with a neural network trained on synthetic data | daily.dev
og:description: A detailed technical experiment trains neural networks to detect lateral movement in corporate networks using entirely synthetic training data generated from a...
og:url: https://daily.dev/posts/looking-for-lateral-movement-with-a-neural-network-trained-on-synthetic-data-f7vvndwzi
og:image: https://api.daily.dev/og/posts/F7VVNdWZI.png
og:image:alt: Looking for lateral movement with a neural network trained on synthetic data
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Looking for lateral movement with a neural network trained on synthetic data

**[Medium](https://daily.dev/sources/medium_js)** · 51 min read · 0 upvotes · 0 comments

## Summary

A detailed technical experiment trains neural networks to detect lateral movement in corporate networks using entirely synthetic training data generated from a 135-line configuration, then evaluates them against the real, labelled Los Alamos National Laboratory authentication logs (1.65 billion events). The best ensemble of six networks trained on six different synthetic worlds surfaces sixteen true attacks within the top 23 ranked windows on a held-out set, versus 161,000 false alarms for a naive threshold counter. The piece walks through architecture search (a small LSTM beats fully connected nets, wide nets, and transformers), why AUC is misleading for rare-event search compared to 'cost of search', two self-corrected conclusions (on coincidences and on ensembles), and an iterative loop where inspecting top false alarms on real data reveals missing phenomena (failure storms, employee relocation, new-machine rates) that get encoded back into the generator config, cutting false alarms from 188 to 7 before the sixteenth attack.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://medium.com/@nick.v.liapin/looking-for-lateral-movement-with-a-neural-network-trained-on-synthetic-data-7e5901362222>

## Questions this post answers

### Can a neural network detect lateral movement attacks without ever training on real attack data?

Yes, a network trained purely on a synthetic corporate network simulation (a 135-line generator config with no real rows) successfully ranked real Los Alamos National Laboratory authentication logs, placing 16 of 16 real red-team attacks within the top 32 of 231,787 windows, versus 161,000 false alarms needed by a naive threshold counter to reach the same recall.

_For teams weighing synthetic-data approaches to security detection, daily.dev surfaces experiments like this one._

### Why is AUC a misleading metric for evaluating rare-event detectors like lateral movement classifiers?

AUC averages ranking quality across all pairs, including millions of easy negatives, so a model can score high AUC while leaving the actual top of the ranked list cluttered with false alarms. On one held-out test, gradient boosting scored a higher AUC (0.918) than a recurrent network (0.909) but required 8,053 false alarms to catch 8 attacks versus 102 for the recurrent network, since boosting only excelled at sorting the hopeless tail nobody reviews.

_Anyone building rare-event detection should track cost-of-search style metrics; daily.dev helps surface this kind of applied evaluation critique._

### Why does training an ensemble of models on merged synthetic datasets perform worse than training separate models on separate datasets?

Merging six differently-configured synthetic worlds into one training set produced 26,297 false alarms before catching the sixteenth attack, roughly 140 times worse than a single-world baseline of 188, because the varied phenomena contradicted each other within one model. Training six separate networks on the six distinct worlds and averaging their rank positions instead reduced false alarms to just 7, since disagreement between diverse models could be resolved by ensembling rank order.

_Developers designing ensemble strategies for anomaly detection can find this kind of concrete before/after comparison on daily.dev._

## Similar posts on daily.dev

- [How a Neural Network Learned Its Own Fraud Rules: A Neuro-Symbolic AI Experiment](https://daily.dev/posts/how-a-neural-network-learned-its-own-fraud-rules-a-neuro-symbolic-ai-experiment-a4siik3dn) · Towards Data Science · 1 upvotes · 0 comments
- [Fine-tune your models with Skills](https://daily.dev/posts/fine-tune-your-models-with-skills-eaqd8sgak) · Event Store · 0 upvotes · 0 comments
- [Anomaly detection on financial data — A classical ML based approach](https://daily.dev/posts/anomaly-detection-on-financial-data-a-classical-ml-based-approach-vnptvnil3) · Medium · 0 upvotes · 0 comments
- [Videos: Humanoid Robot Martial Arts, Perseverance, More](https://daily.dev/posts/videos-humanoid-robot-martial-arts-perseverance-more-viyh3mebb) · IEEE Spectrum · 1 upvotes · 0 comments

---

Tags: [#security](https://daily.dev/tags/security), [#machine-learning](https://daily.dev/tags/machine-learning), [#neural-networks](https://daily.dev/tags/neural-networks)

[View this post on daily.dev](https://daily.dev/posts/looking-for-lateral-movement-with-a-neural-network-trained-on-synthetic-data-f7vvndwzi)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"TechArticle","headline":"Looking for lateral movement with a neural network trained on synthetic data","url":"https://daily.dev/posts/looking-for-lateral-movement-with-a-neural-network-trained-on-synthetic-data-f7vvndwzi","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/looking-for-lateral-movement-with-a-neural-network-trained-on-synthetic-data-f7vvndwzi"},"datePublished":"2026-09-17T04:16:28.617Z","dateModified":"2026-09-17T04:41:39.013Z","description":"A detailed technical experiment trains neural networks to detect lateral movement in corporate networks using entirely synthetic training data generated from a...","image":"https://media.daily.dev/image/fetch/s--3UU9YYqe--/f_auto,pg_1,q_auto/https://miro.medium.com/v2/da:true/resize:fit:1200/1%2AGFhgDynMw8DRa5QcrJRo6g.gif?_a=BAMAMiZY0","thumbnailUrl":"https://media.daily.dev/image/fetch/s--3UU9YYqe--/f_auto,pg_1,q_auto/https://miro.medium.com/v2/da:true/resize:fit:1200/1%2AGFhgDynMw8DRa5QcrJRo6g.gif?_a=BAMAMiZY0","isAccessibleForFree":true,"articleSection":"Medium","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Medium","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/medium","url":"https://daily.dev/sources/medium_js"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/looking-for-lateral-movement-with-a-neural-network-trained-on-synthetic-data-f7vvndwzi","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"security,machine-learning,neural-networks","timeRequired":"PT51M"}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Medium","item":"https://daily.dev/sources/medium_js"},{"@type":"ListItem","position":3,"name":"Looking for lateral movement with a neural network trained on synthetic data"}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/looking-for-lateral-movement-with-a-neural-network-trained-on-synthetic-data-f7vvndwzi#faq","mainEntity":[{"@type":"Question","name":"Can a neural network detect lateral movement attacks without ever training on real attack data?","acceptedAnswer":{"@type":"Answer","text":"Yes, a network trained purely on a synthetic corporate network simulation (a 135-line generator config with no real rows) successfully ranked real Los Alamos National Laboratory authentication logs, placing 16 of 16 real red-team attacks within the top 32 of 231,787 windows, versus 161,000 false alarms needed by a naive threshold counter to reach the same recall. For teams weighing synthetic-data approaches to security detection, daily.dev surfaces experiments like this one."}},{"@type":"Question","name":"Why is AUC a misleading metric for evaluating rare-event detectors like lateral movement classifiers?","acceptedAnswer":{"@type":"Answer","text":"AUC averages ranking quality across all pairs, including millions of easy negatives, so a model can score high AUC while leaving the actual top of the ranked list cluttered with false alarms. On one held-out test, gradient boosting scored a higher AUC (0.918) than a recurrent network (0.909) but required 8,053 false alarms to catch 8 attacks versus 102 for the recurrent network, since boosting only excelled at sorting the hopeless tail nobody reviews. Anyone building rare-event detection should track cost-of-search style metrics; daily.dev helps surface this kind of applied evaluation critique."}},{"@type":"Question","name":"Why does training an ensemble of models on merged synthetic datasets perform worse than training separate models on separate datasets?","acceptedAnswer":{"@type":"Answer","text":"Merging six differently-configured synthetic worlds into one training set produced 26,297 false alarms before catching the sixteenth attack, roughly 140 times worse than a single-world baseline of 188, because the varied phenomena contradicted each other within one model. Training six separate networks on the six distinct worlds and averaging their rank positions instead reduced false alarms to just 7, since disagreement between diverse models could be resolved by ensembling rank order. Developers designing ensemble strategies for anomaly detection can find this kind of concrete before/after comparison on daily.dev."}}]}
```

