Home › Topics › Detecting Traffic Anomalies › NDR for Generalists

Network Detection and Response Concepts for Generalists


At some point someone — a vendor, an auditor, a manager who read an article on a plane — will ask whether you have "network detection and response." You will nod thoughtfully. Then you will go back to your desk and search for what it actually means. You will find a wall of product pages that all promise to find the threats you cannot see, usually with a glowing padlock picture. This guide is the plain-language answer you were looking for.


Network detection and response (NDR) is a category of tooling that watches network traffic, learns what normal looks like, flags departures, and helps you act on them. Strip away the marketing and it is the industrialized version of exactly what the other guides in this topic teach by hand. That includes baselines, beacon detection, scan detection, and triage. That is good news, because it means you already understand most of what the products do. You can make a sober judgment about what a product would add for your particular network and team.


By the end of this guide you will be able to explain what these products do step by step. You can draw an honest line between what you can approximate yourself and what genuinely needs a product. You can frame the budget question in terms a small team can act on. And you can enter a purchase conversation with a list of questions that separates substance from glow.


In this guide

  • What "detection and response" means in plain words
  • What detection products do, step by step
  • What you can approximate yourself
  • What genuinely needs a product
  • DIY versus product: a capability comparison
  • The honest budget framing for a small team
  • Questions to ask before buying
  • Feed your alerting; do not replace it
  • Where to go next

What "detection and response" means in plain words


Take the phrase apart. Network means the data source is traffic on the wire — flow records, connection metadata, sometimes full packet captures — rather than logs from inside a host. Detection means the tool's job is to notice things worth a human's attention, mostly by comparing current behavior to learned behavior. Response means the tool helps you do something about it. At minimum, it packages evidence and context. At most, it can isolate a host or push a block to a firewall on your behalf.


The idea is a cousin of the older intrusion detection system, which matches traffic against a library of known-bad patterns, the way antivirus matches files against known signatures. Signature matching is precise but only catches what someone has already described. Detection products of the newer sort lean instead on behavior. They do not need to know what a given piece of unwanted software looks like. They only need to know that this host has never talked to that address every ninety seconds before. The two approaches complement each other, and many products do both; the intrusion detection guides cover the signature side.


One more definition, because the vendor will use it: telemetry is just the word for the measurements a tool collects. Flow records are telemetry. So are DNS query logs and connection summaries. When a product page says it "ingests rich telemetry," it means it reads a lot of the same data you already have or could have.


What detection products do, step by step


Underneath every product in this category is the same pipeline. Understanding the seven stages lets you evaluate any tool by asking how well it does each one, rather than being carried along by the demo.


  • 1. Ingest flows and metadata. The product collects flow records exported by your routers, switches, and firewalls. Or it watches a mirrored copy of the traffic itself and generates its own records. Some also pull DNS logs, proxy logs, and directory logs to know who is who. If it cannot see a segment, it cannot detect anything on that segment; visibility is the foundation everything else stands on.
  • 2. Learn a baseline per host. For each device it sees, the product builds a picture of normal: how much it sends and receives, when, to whom, on which ports. This is the same idea as the baseline you would build yourself, done automatically for every host, and usually broken out by hour and day of week.
  • 3. Model relationships. Beyond single hosts, the product keeps a map of who normally talks to whom. It maps which workstations use which servers, which servers replicate to which peers, and which hosts reach the internet at all. New edges on that map are one of the strongest signals it has.
  • 4. Score deviations. Each new observation is compared to the baseline and the relationship map, and the size of the departure is turned into a score. A host slightly above its usual volume scores low; a host talking to a brand-new external address on a schedule at 3 AM scores high. Scoring is where products differ most and where vendors are least specific.
  • 5. Enrich with context. A raw score is enriched with everything the product knows. That includes the host's role from the directory and the destination's reputation from a threat-intelligence feed. It also includes the user logged in at the time and whether peers behave the same way. This stage is what turns "unusual" into "unusual and unexplained."
  • 6. Alert. Findings that clear a threshold become alerts, ideally grouped so that one incident produces one alert rather than fifty. Better products present a narrative — "this host scanned the subnet, then connected to three responders, then began beaconing" — rather than three unrelated tickets.
  • 7. Respond (sometimes). Some products stop at alerting. Others can quarantine a host through your switches, push a block to your firewall, or open a ticket with the evidence attached. Automated response is powerful and dangerous in equal measure. Most small teams keep it in "recommend, do not act" mode for a long time, and that is sensible.

Look at that list again and notice how familiar it is. Stages one through four are what the beaconing and scanning guides in this topic walk you through with a flow table and a query. Stage five is the triage question set. Stage six is your alerting pipeline. What the product adds is doing all of it, for every host, continuously, with better models and more context than you can assemble by hand.


What you can approximate yourself


A good chunk, honestly. A generalist needs flow export from the network gear, a place to store the records, and a willingness to write a few queries. With those, you can cover the detections that matter most in a small or mid-sized network:


  • Volume anomalies. A host far above or below its own typical bytes for the hour. This is the easiest rule to write and the one your existing monitoring may already do at the interface level.
  • New external destinations. A host contacting an address that neither it nor any peer has contacted before. One query against a "seen before" table.
  • Beaconing. Regular intervals and consistent sizes to one destination, measured over a day — the beaconing guide gives you the sketch.
  • Internal scanning. Distinct destinations per source per minute, distinct ports per pair, deny-log fan-out — the scanning guide gives you those.
  • Off-hours activity. Any of the above, weighted higher outside the host's normal working window.
  • Role violations. A printer speaking a management protocol outbound; a workstation serving file shares. A short list of "this role should never do that" rules, checked against flows.
  • Top talkers and hogs. Who is using the link, and whether that is a legitimate or a suspicious hog — a question that is as much operational as it is security.

That is not a toy list. It is, roughly, the set of detections that catches the loud and the medium-loud problems. For a network of a few hundred hosts it can be run from a database and a scheduled job. The catch is not capability; it is time. Every rule needs writing, tuning, and reviewing, and every alert it produces lands on the same generalist who has printers to fix. Be honest about that when you weigh the next section.


What genuinely needs a product


There is a line past which the do-it-yourself approach stops being a good use of a small team's hours. And it is worth naming exactly where it is.


Scale

A few hundred hosts and a handful of subnets fit in a database and a query. Thousands of hosts, dozens of sites, cloud workloads that appear and vanish hourly, and millions of flow records a minute do not fit. At least, fitting them requires building a data platform that becomes its own full-time job. Products are engineered for that scale; your scheduled job is not.


Models

Your queries use fixed thresholds you chose. Products use statistical models that adapt per host, per hour, and per relationship, and that can score a departure across several dimensions at once. You can approximate a lot with a coefficient of variation and a "seen before" table. But a well-built model finds subtler things that a handful of rules will not. It might find a slow change in a host's peer set over weeks. Or it might find a destination that is new to the whole industry rather than just to you.


Threat intelligence

Enrichment with what the wider world knows requires feeds you cannot easily assemble or keep current yourself. Those feeds might tell you that this address range has been seen hosting controllers, or that this domain was registered yesterday. Products bundle and maintain them. This is often the single most valuable thing a product adds for a generalist. It converts "new to us" into "new to us and known to be bad elsewhere."


Expert tuning and narrative

A product ships with detections written and tuned by people who do nothing else, and it groups related findings into a story. Your rules are as good as the hours you can give them. Some products also include a service tier where the vendor's analysts watch the alerts for you. For a very small team, that can be the difference between having detection and merely owning it.


Response integration and retention

You can build ways to isolate a host through a switch API, push a block to the firewall, or open a ticket with evidence attached. You can also build a way to keep months of searchable history for the investigation that comes later. Each is possible, and each is a project. Products have already built them.


DIY versus product: a capability comparison


CapabilityFlow data + rules + baselines (DIY)Detection product
Visibility into who talks to whomYes, wherever you export flowsYes, plus its own sensors where you cannot export
Per-host baselinesManual or scripted, coarse (daily or hourly totals)Automatic, per host, per hour, per relationship
Beacon and scan detectionYes, with the queries in this topicYes, with more resilience to jitter and evasion
New-destination detectionYes, simple and effectiveYes, enriched with destination reputation
Multi-dimension scoringRough (count how many rules fired)Built in, weighted, adaptive
Threat intelligence enrichmentLimited or noneBundled and maintained
Grouping findings into a narrativeYou do it by hand in the ticketOften automatic
Automated responseOnly what you script yourselfAvailable, usually optional
Long-term searchable historyAs long as your database and disk allowIncluded, subject to licensed retention
Tuning effortYours, ongoingShared with the vendor; still yours for local context
CostHours, storage, and attentionLicense, deployment, and still some hours
Best fitSmall networks, generalist teams, learning the craftLarger or higher-risk networks, teams with a security remit

The last row is the honest summary. Neither column is "correct." A small network with a generalist who has read this topic and runs a dozen tuned rules is in a genuinely good position. A larger network with regulatory obligations and no one who can give the rules regular attention is not, and that is the case for a product.


The honest budget framing for a small team


The trap is comparing a product's license to zero. The do-it-yourself path is not free; it costs your hours, and your hours are the scarcest thing in a small team. So frame the decision as a comparison between two real costs.


First, estimate the hours. Building the flow pipeline and the first rules is a one-time effort measured in days. Keeping them tuned, reviewing what fires, and investigating candidates is a recurring effort measured in hours per week. It does not shrink to zero, because networks change. Write the weekly number down honestly. If the answer is "an hour or two a week and it fits," the do-it-yourself path is probably right for now. Suppose the answer is "it would be several hours a week that I do not have, so in practice it would not get done." Then you do not have detection; you have a script. In that case, the product conversation is worth having.


Second, weigh what is at stake. Detection buys time: it turns a problem you would have discovered when the ransom note appeared into one you discover when the beacon starts. How much is that time worth to your organization? A small firm with modest data and good backups may reasonably decide that flows-plus-rules is proportionate. A firm holding sensitive records, subject to regulation, or one bad week away from real damage, has a different answer.


Third, stage it. For most small teams, the path starts with flow export and a top-talkers view, because it pays for itself operationally on day one. The anomaly rules from this topic come second, run in log-only mode and tuned monthly. Add a product when — and only when — the candidate queue is reliably larger than the hours available to review it, or when the stakes change. Buying the product first, before the flows and the habits exist, tends to produce an expensive dashboard nobody reads.


Tip: if you do buy, keep your own simple rules running alongside the product for the first months. They are your independent check: if the product misses something your beacon query catches, you have learned something important about the product. If it catches everything you do and more, you can retire the rules with confidence rather than on faith.


Questions to ask before buying


Product conversations go better when you bring the questions. These are generic, apply to any vendor, and map directly to the seven pipeline stages above.


  • 1. What does it need to see, and how? Flow export from your existing gear, a mirror port, its own appliance, an agent on every host, or some mix? Which of your segments — branch sites, wireless, cloud — would it not see with the proposed deployment?
  • 2. How long does the baseline take to learn, and what happens during that time? Every product has a learning period during which alerts are unreliable. Ask how long, and how you will know it has settled.
  • 3. How are alerts scored, and can I see why? A score with no explanation is a verdict machine. You want a product that shows its work: which baseline, which deviation, which enrichment.
  • 4. How do I allowlist by identity and role, not just address? Your monitoring server, backup server, and vulnerability scanner will trip everything on day one. Ask exactly how those exceptions are expressed and whether they survive an address change.
  • 5. What does it do with my existing tools? Can it send alerts into the ticketing, chat, or on-call system you already use? Can it read your directory for host roles and your firewall logs for context?
  • 6. What does response actually mean here? Recommendation only, or action? Which actions, through which of your devices, with what approval step? Can automatic action be turned off entirely at first?
  • 7. How much of my time will it need each week, honestly? Ask for the number. A vendor who says "none" has not deployed it in a small team.
  • 8. What does retention cost, and what happens to my data if I leave? Investigations look backward. Know how far back you can look, and whether you can take your history with you.

Feed your alerting; do not replace it


The most common deployment mistake is standing a detection product up as a second, separate alerting universe. It has its own console, its own notifications, its own severity scale, and its own on-call expectations. Two universes means two places to look, two things to ignore, and a guarantee that the one nobody opens on Friday afternoon is the one that mattered.


Treat detection tooling as one more source flowing into the alerting pipeline you already run. Its findings should arrive in the same queue as a down interface or a full disk, with a severity on the same scale. They should be grouped and de-duplicated by the same logic, and page the same rotation. That is partly a technical integration. Most products can send to a webhook, an email, or a ticketing system. It is mostly a discipline: one queue, one severity scale, one set of eyes.


This is also where the monitoring you already have earns unexpected credit. The flow records and interface graphs that tell you who is filling the link are the same records that reveal a beacon. The auto-discovery that maps your network is the same data a detection product uses to know what a host is. Treating operations data and security data as one dataset seen through two lenses is not a slogan; it is the practical reason a generalist can do this work at all.


Where to go next


You can now explain network detection and response without the padlock: ingest flows, learn baselines, model relationships, score deviations, enrich, alert, and sometimes respond. You know that a generalist with flow data and a few tuned rules can approximate a meaningful share of it. Scale, models, threat intelligence, expert tuning, and response integration are where a product earns its license. The honest budget question is hours against hours, not license against zero.


Whichever path you take, the underlying craft is the same. The rest of the Detecting Traffic Anomalies topic teaches it. The guide to what counts as an anomaly gives you the dimensions and the triage questions. The beaconing and scanning guides give you the detections. And tuning out false positives keeps the whole thing readable — which matters just as much with a product as without one.


Put this into practice with SysPeep Monitor

Detection tooling, bought or built, stands on the same foundation: knowing what is on your network and how it normally behaves. SysPeep Monitor is not a detection-and-response product, but it gives you that foundation. It offers auto-discovered devices and interfaces, health and traffic metrics collected over SNMP, WMI, and SSH, and alert rules built to keep the noise down. All of this comes from one Windows service and one web console. The free trial is on the download page.

Download Free TrialLearn About SysPeep