It is Tuesday, 9:40 a.m. The sectoral CSIRT asks for clarification of the early warning: which machines in the internal network did the infected file server talk to over the last two weeks? The firewall logs show traffic at the network perimeter. EDR has a full history of the workstations, but not of the printers, IP cameras or controllers on the shop floor. Without a history of traffic inside the network, the honest answer is “we don’t know” — and that is a weak starting point for a report that must be supplemented with an impact assessment in less than three days.
Flow-based network traffic analysis answers exactly that question: who talked to whom, when and how much. It does not replace the firewall, EDR or packet capture; it fills the gap between them. In this article we describe what can be detected from traffic metadata alone, where the method ends and how its data translates into obligations under the NIS2 Directive and the amended Polish Act on the National Cybersecurity System. Progress Flowmon serves as the example tool, but the criteria apply to any platform of this class.
What is network traffic analysis, and how do flows differ from packets?
It is the study of communication in a network based on a record of it — either full packets or their summaries, i.e. flows. Packet capture records everything, including the content, at one point in the network. That is how Wireshark works, which we describe in the article on Wireshark and packet capture. The method is precise, but recording the whole network for weeks quickly becomes unattainable.
A flow is a summary. The IPFIX protocol specification (RFC 7011, September 2013) defines it as a set of packets passing an observation point during a certain time interval that share common properties, for example addresses, ports and protocol. A single flow record says which devices talked, for how long, over which port and how much data they exchanged. It does not say what was inside.
This difference determines the use. Packets serve for detailed analysis of a specific event. Flows serve for knowledge of the whole network over a long period: they let you see patterns, deviations from normal and a history you can return to weeks later. In practice the two methods complement each other, and flow analysis points to where it is worth reaching for packets.
A flow in one sentence: a “who, with whom, when, over which port and how much data” record for every conversation in the network, without the content of that conversation. Cheap to store, so it can reach far back.
What do firewalls and EDR miss inside the network?
Traffic that does not cross the network perimeter, and devices without an agent. An edge firewall controls connections to the internet and between separated zones. A connection from a workstation to a file server in the same segment does not touch it at all. This is “east–west” traffic, which the attacker uses to move around after getting into the network.
EDR sees in detail what happens on a device with an installed agent: processes, files, connections. But the agent does not go onto printers, IP cameras, VoIP phones, network devices, embedded systems or industrial controllers. We compare the scope of these tool classes in the article EDR vs XDR vs NDR.
This gap is filled by the category called network detection and response, or NDR, which we define in more detail in the article on NDR and network visibility. Progress positions Flowmon exactly here: between perimeter protection and endpoint protection. Traffic analysis does not need an agent on the device, because it observes the device’s communication from the outside.
The less segmented the network, the more lateral traffic bypasses every control point — we describe how to limit it in the article on network microsegmentation.
Where do you get flow data from: routers, switches or probes?
First from devices already in the network, and use probes where that data is not enough. Routers, switches, firewalls and load balancers can usually export flow records. According to the vendor, the Flowmon collector accepts NetFlow versions 5 and 9, IPFIX, sFlow, jFlow and other flow formats described in RFCs. It also handles flow logs from the AWS, Azure and Google Cloud clouds.
Before turning on export, you have to check one thing: sampling. Many devices, especially at high throughput, export only every n-th packet, and sFlow is sample-based by design. Sampling is enough for link utilisation statistics, but it loses short connections — exactly the ones that reveal scanning and login attempts. What is not in the export, the collector can no longer reconstruct.
A probe is a separate device that observes traffic passively — through a switch SPAN port or through a TAP — and creates the records itself. The Flowmon probe comes in hardware, virtual and cloud versions, for links from 10 Mb/s to 100 Gb/s. Beyond layer 2–4 data, it adds application-layer information in an IPFIX extension, such as host names, URLs and DNS queries, as well as performance metrics: server response time, delay, jitter and retransmissions.
A sensible setup combines both methods: export from devices gives a broad picture at low cost, and a probe in front of key servers gives complete data where it is most valuable.
Which threats can be detected from traffic metadata alone?
Those that change the pattern of communication, even if they do not change its content. Network reconnaissance looks like one device connecting to many addresses or ports in a short time. Lateral movement shows up as new SMB or RDP connections between workstations that previously did not talk to each other. Exfiltration is an unusual volume of outbound data, often at an unusual time or to a new destination.
Connectivity to the attacker’s command server often takes the form of regular, small connections at fixed intervals. Misuse of auxiliary protocols is another class. Progress lists analysis of DNS, DHCP, ICMP and SMTP as a way to detect exfiltration, reconnaissance and lateral movement.
Flowmon does this in its Anomaly Detection System module. The vendor claims many AI-based methods and algorithms combining behavioural analysis, adaptive baselines, address reputation and signature detection. It maps events onto the MITRE ATT&CK framework and can compare traffic against indicators of compromise from the MISP information-sharing community. An optional IDS probe based on the Suricata engine adds signature detection on packets.
Metadata also serves for threat hunting: the hypothesis “who outside the usual group connects to this server” can be checked against flow history, without installing anything on the server. We describe the method in the article on threat hunting in practice.
What can you see in encrypted traffic without decrypting it?
More than people usually assume, though not the content. Encryption hides what the conversation contains, but not the fact that it happened, between whom, for how long and at what volume. The patterns described above — scanning, lateral movement, regular command connections — work the same way for encrypted traffic.
The TLS protocol itself also leaves traces. The Flowmon probe enriches records with information from the TLS negotiation. Among its uses, the vendor lists flagging out-of-date and policy-non-compliant certificates, weak encryption and old TLS versions that may contain vulnerabilities.
This matters not only for detecting attacks. Article 21(2)(h) of the NIS2 Directive requires policies and procedures on the use of cryptography and encryption. An inventory of the TLS versions and certificates actually in use, built from traffic observation, shows whether the policy works in practice and not just on paper.
The boundary is firm: metadata will not tell you whether a specific file was sent in an encrypted connection.
How long should flow data be kept, and how much space does it take?
Long enough to reach back to the likely moment of the attacker’s first entry, not just to the day of detection. Incidents are rarely detected on the day they began. A history covering only the last few days lets you describe the symptoms, but not the path the attacker took through the network.
The data volume depends on the number of flows, not on link throughput. A network with many short connections, for example a large number of DNS queries, generates more records than a network with a few large transfers. That is why sizing is best based on a measurement from a pilot rather than on a catalogue ratio.
Progress states a capacity of 0.5 to 192 TB per appliance for the Flowmon collector and data retention from months to years, depending on the chosen capacity. The vendor stresses that the collector neither aggregates nor samples data. Larger networks are served by a distributed architecture with a central node that gathers data from branch offices.
The records contain addresses that can be attributed to employees, so the retention period and access scope should be documented as for other security logs.
How does traffic analysis support reporting obligations under NIS2 and the Polish cybersecurity act?
It provides the facts you have to state in successive notifications, within deadlines in which you otherwise would not have them. Article 23(4) of the NIS2 Directive provides for an early warning within 24 hours of becoming aware of a significant incident. Within 72 hours an incident notification must follow, with an initial assessment of severity and impact and, where available, indicators of compromise. The final report, with the type of threat or root cause, is due no later than one month after the notification.
The Polish transposition counts these deadlines from the detection of the incident. The amended Act on the National Cybersecurity System (Journal of Laws 2026, item 252, in force since 3 April 2026), in Article 11(1), points 4 and 4a, requires an early warning within 24 hours and notification of a serious incident within 72 hours of its detection, to the competent sectoral CSIRT. Under Article 33(1) of the amending act, essential and important entities implement the obligations of Chapter 3 of the act within 12 months of its entry into force.
The Directive defines incident handling as actions aimed, among other things, at detecting, analysing and containing an incident (Article 6, point 8). Flow history answers the questions that arise at every stage: since when the unusual communication has been going on, which devices are involved, whether data left the network and in what quantity. Without it, the initial impact assessment rests on assumptions.
Reporting timeline and network data: 24 h — whether the traffic points to deliberate action and where it came from; 72 h — which segments and devices are affected, whether there was a transfer outside; 1 month — the attacker’s path from the first connection, the root cause, the effectiveness of containment. Every stage needs history reaching back before the day of detection.
The tool does not ensure compliance on its own, however. NIS2 requires an incident handling process, assigned roles and documentation. Traffic analysis is one of the data sources for that process, and we review all the requirements as part of NIS2 compliance implementation. We have collected our other articles on the Directive in the NIS2 knowledge base.
Where does flow analysis end, and when do you need packet capture?
Where the question concerns content rather than the fact of communication. A flow will tell you that a few gigabytes left the server for an unknown address. It will not tell you which files they were. To establish the scope of a leak or reconstruct a specific command, you need packet capture or data from the device itself.
The second boundary is sampling and unobserved places. If the switch export is sampled or traffic bypasses measurement points, the history will be incomplete. The third boundary is processes on the device: traffic analysis will not show which program opened the connection. That is EDR territory.
Flowmon softens the first boundary by triggering full packet capture when an event is detected; according to the vendor, a rolling buffer also covers traffic from before the attack began.
In a post-incident investigation, flows show the direction, while packets and device data provide the evidence. We describe the order of securing evidence in the article on digital forensics after a cyberattack.
How do you connect NDR with SIEM and the SOC so that alerts do not drown?
Through three connections: event export, identity enrichment and assigning an owner. The first is technical. Flowmon forwards events to log management systems and SIEM via syslog, SNMP, e-mail, REST API or custom scripts. This way a network alert lands next to alerts from EDR and system logs.
The second connection turns an IP address into a person. Flowmon can correlate events with logs from authentication systems, for example Active Directory, LDAP or Cisco ISE. The analyst then sees a user or host name instead of a bare address. With dynamic addressing, that is the difference between a quick decision and an hour spent working out who had a given address.
The third connection is organisational and the one most often skipped. Someone has to read the alert, assess it and decide on the response — during the hours when the attacker is active. Automated response helps, because Flowmon can, for example, isolate an address through integration with a network access control system. It does not, however, replace human judgement for ambiguous events.
Without your own on-call team, network alerts are worth feeding into a SOC service — otherwise detection ends on a dashboard that nobody watches at night.
How do you choose a traffic analysis tool?
Based on requirements you can test in a pilot, not on the feature list in vendor materials. The table below turns each requirement into a question for the vendor and a way to verify it in your own network.
| Requirement | Question for the vendor | How to check in a pilot |
|---|---|---|
| Data sources | Which formats does it accept (NetFlow v5/v9, IPFIX, sFlow, cloud logs)? | connect exports from your own devices and cloud, compare record counts |
| Completeness | Does the collector aggregate or sample data? | compare records from a probe and from a sampled export for the same link |
| Retention | How many months of history fit at our number of flows? | measure flows per second for a week and calculate capacity |
| Application layer and TLS | Do records include DNS, host names and TLS metadata? | check whether the report flags old TLS versions in our network |
| Detection | Which detection methods, and are events mapped to MITRE ATT&CK? | run a controlled scan of a segment and measure time to alert |
| Identity | Does it correlate addresses with AD, LDAP or NAC logs? | check whether the alert shows a user instead of an address |
| Integrations | How does it export events to SIEM and SOAR? | send an event to your own SIEM and check the full set of fields |
| Packet capture | Can it trigger capture after an event, including traffic before it? | trigger a test event and read the capture |
| Deployment | Are hardware, virtual and cloud versions available? | run the virtual version in the target environment |
The order of the rows is not accidental. Completeness and retention decide whether the tool will answer the questions in a post-incident report. Detection and integrations decide whether the incident will be detected earlier.
Where should a mid-sized company start the deployment?
With a map of the points traffic flows through, and with the data the network already produces. The first step is to identify observation points: the network core, the link to the server room, the internet exit and cloud environments. The second is to turn on flow export on existing devices and check that it is not sampled.
The third step is a probe where device export is not enough, usually in front of the most important servers. The fourth is a period of learning normal traffic and tuning thresholds before alerts reach the response team. The fifth is a test: a controlled scan or a simulation of lateral movement, to check whether and when an alert appears.
Finally, the result has to be wired into the incident handling procedure: who receives the alert, who decides on isolation and who prepares the data for the 24-hour notification.
Pilot checklist: observation points identified; export without sampling or with sampling documented; number of flows measured and retention calculated; scan detection tested; alert visible in SIEM with a user name; alert owner assigned.
We run Progress Flowmon pilots and deployments together with configuring export on existing devices and bringing events into monitoring. Ongoing infrastructure supervision is part of our IT infrastructure monitoring service. You will also find the terms from this article in our glossary: NDR and network monitoring.
Sources cited in the text: RFC 7011 “Specification of the IP Flow Information Export (IPFIX) Protocol” (September 2013); Directive (EU) 2022/2555 (NIS2), Article 6(8), Article 21(2), Article 23(4); Polish Act of 23 January 2026 amending the Act on the National Cybersecurity System (Journal of Laws 2026, item 252), Article 11(1) points 4 and 4a, Article 33(1); Progress Flowmon product pages (progress.com, as of September 2026).
Frequently asked questions
How does flow analysis differ from packet capture?
Packet capture records the full content of traffic at one point in the network; it is precise but expensive to store. A flow, according to the IPFIX specification (RFC 7011, September 2013), is a set of packets with common properties, such as addresses and ports, passing an observation point during a certain time interval. A flow record says who talked to whom, when and how much data they exchanged, without the content. That is why it can be kept for months for the whole network.
Does traffic analysis work when traffic is encrypted?
Partly. The connection metadata remains visible: addresses, ports, volume, timing, and with TLS also the protocol version, cipher strength and certificate validity. Progress Flowmon uses this, among other things, to flag out-of-date certificates and old TLS versions. Flow analysis cannot read the content of encrypted communication.
Is a traffic analysis tool enough to comply with NIS2?
No. Article 21(2) of the NIS2 Directive lists, among other measures, incident handling and business continuity, and Article 23(4) sets reporting deadlines: 24 hours for an early warning, 72 hours for an incident notification and one month for the final report. Traffic analysis provides data for detecting and describing an incident, but compliance requires a process, roles and documentation.
Where do you get flow data if you do not buy probes?
From devices already in the network. Routers, switches, firewalls and load balancers usually export NetFlow, IPFIX or sFlow, and public clouds provide their own flow logs. You do need to check whether the export is sampled, because sampling loses short connections that the collector can no longer reconstruct.
How long should flow data be retained?
Long enough to reach back to the likely moment of the attacker's first entry, not just to the day of detection. In practice, sizing starts with measuring the number of flows during a pilot. Progress states a capacity of 0.5 to 192 TB per appliance for the Flowmon collector and data retention from months to years.
