← all systems · see it in The System
Intrusion detection system
A device or software application that monitors a network or systems for malicious activity or policy violations, reporting it to an administrator or collecting it centrally (src:wp:intrusion-detection-system); intrusion detection is 'the process of monitoring the events occurring in a computer system or network and analyzing them for signs of possible incidents' (src:nist-sp800-94). Walked as the technology and the practice around it: sensors and agents, the decoding and detection engines (signature, anomaly, stateful protocol analysis), alerts and records, the analysts who validate them, and the site policy that says what counts as an intrusion.
Properties published scienceour proposal
Sourced fields, and our own readings.
| aliases | Intrusion detection |
| born at | 1980 (Anderson's study); 1986 (Denning's model, the basis of many systems) |
| born precision | years |
| boundary | Inside: the sensors and agents that watch (network taps, packet capture, host agents, audit-trail readers, wireless sensors), the decoding of what they see into events, the detection engines (signatures, profiles of normal behaviour, protocol-state models), the alerts, logs and reports, the management servers, databases and consoles, and the people who tune it and validate its alarms, with the site policy that says what counts as an intrusion. Outside: the monitored hosts and networks themselves; the intruders and the benign users whose activity it reads; the firewall and access control that block (the edge is a contention: an intrusion prevention system 'performs access control like an application layer firewall'); the SIEM and security operations centre that collect its output; incident response beyond the alert. |
| nested level | A security technology (rung 12) built on computer networks and operating systems, one of the detection controls of computer security and information security; its alerts feed SIEMs and security operations centres, where people read them. |
| scale | technological |
Measures (13) published science
| connection-attempt timer (Bro) | 5 minutes |
| characteristic clock: alert presentation rate in a cyber vigilance task | 8 (slow) or 16 (fast) events per minute |
| wireless channel scanning | each channel a few times per second visits per second |
| training period for an anomaly profile | days, sometimes weeks |
| vigilance decrement in Mackworth's clock test | correct detections down 10% in the first 30 min of a 2-h task percent |
| audit records per day at a small installation | about 1,000,000 (against one or two intrusion attempts a day) records per day |
| Bayesian detection rate at a false alarm rate of 1e-5 | about 66% with detection rate 1.0; about 58% with detection rate 0.7 share of alarms that are true |
| password-failure threshold (Denning's example) | more than 10 in a brief period failures |
| share of alarms analysts call false positives | 99% (one analyst's estimate, 7 SOCs, n = 21 interviews) percent |
| Snort active users | over 300,000 users |
| Kitsune throughput on a Raspberry Pi | about 1,000 packets/s with one autoencoder; about 5,400 with 35 packets per second |
| cost of Target's malware-detection tool (2013) | $1.6m US dollars |
| web share of border bandwidth in an example profile | 13% on workdays percent |
Open contentions (2) open contention
Where sources or readings disagree. Both stay on the map.
- C-1482 Is machine-learning anomaly detection ready for real networks?
booked: The map books it as a working kind that real sites seldom use, and keeps the view that it stays mostly a research tool.
the other reading: anomaly detection by ML is fundamentally harder here than elsewhere (cost of errors, semantic gap, diversity, evaluation) and stays mostly in research; it is practical now - C-1496 Are 99% of security alarms really false?
booked: The map books the analysts' own count, and keeps the finding that most such alarms are real events on harmless activity.
the other reading: most alarms are true alarms on legitimate activity (benign triggers), so the tools are less unreliable than the 99% says; false alarms in the strict sense (no security event) dominate
Partners and parts published science
Built from / builds (14)
- Computer security
- Human (Homo sapiens)
- Transaction log
- Sensors and agents
- Packet capture (libpcap, taps, SPAN ports)
- Decoder and event engine
- Signatures and rules
- Profiles of normal behaviour
- Management servers, database servers and consoles
- Malware evasion against machine-learning detection
- Overload, crash and subterfuge attacks on the monitor
- Intruders against intrusion detection
- Adversarial examples against machine-learning NIDS
- Botmasters against defenders for control of the bots
Sits inside / holds (4)
Trades with (valence) (20)
- Malware via sys:written-record · Intrusion detection system mixed
- Human (Homo sapiens) via sys:human-society · Intrusion detection system mixed
- Operating system via sys:written-record · Intrusion detection system gain
- Machine learning via sys:software · Intrusion detection system mixed
- Cryptography via sys:electromagnetic-field · Intrusion detection system loss
- Computer security via sys:electromagnetic-field · Intrusion detection system gain
- Access control via sys:human-society · Intrusion detection system gain
- Security operations center (SOC) via sys:electromagnetic-field · Intrusion detection system gain
- Information security via sys:written-record · Intrusion detection system gain
- Firewall (computing) via sys:electromagnetic-field · Intrusion detection system gain
- Computer network via sys:electromagnetic-field · Intrusion detection system gain
- Intruder (burglar and trespasser) via sys:electromagnetic-field · Intrusion detection system mixed
- Security information and event management (SIEM) via sys:written-record · Intrusion detection system gain
- Honeypots (decoys) via sys:electromagnetic-field · Intrusion detection system gain
- Cyber threat intelligence via sys:written-record · Intrusion detection system gain
- Incident response via sys:electromagnetic-field · Intrusion detection system gain
- Threat actor via sys:computer-network · Intrusion detection system gain
- Botnet via sys:computer-network · Intrusion detection system gain
- Denial-of-service attack via sys:internet · Intrusion detection system loss
- Physical security via sys:written-record · Intrusion detection system gain
expressed by (23)
- Generation/Synthesis
- Recombination/Composition
- Templating / Standardization
- Quality Improvement/Refinement
- Interoperability/Transduction (Bridging)
- Encapsulation & Modularity (Abstraction)
- Internal Coordination/Regulation
- Sensing/Scanning
- Mapping/Modeling
- Uncertainty Reduction/Learning
- Multi-Scale Data Fusion / Triangulation
- Trail-Marking / Externalized Memory (Stigmergy)
- Surprise / Change-Point Detection
- Frontier Outposts / Sensing Extension
- Social Learning / Public Information Use
- Quick Harm Avoidance
- Reactive Response
- Technique-Based Identification (a.k.a. Technique-Based Detection)
- Self-Monitoring
- Validation of Effectiveness
- Redundancy & Diversity
- Triage & Prioritized Allocation
- Deterrence
floor filled by (20)
- Computer security
- Human (Homo sapiens)
- Operating system
- Machine learning
- Security operations center (SOC)
- Information security
- Computer network
- Transaction log
- Intruder (burglar and trespasser)
- Sensors and agents
- Packet capture (libpcap, taps, SPAN ports)
- Decoder and event engine
- Signatures and rules
- Profiles of normal behaviour
- Management servers, database servers and consoles
- Security information and event management (SIEM)
- Honeypots (decoys)
- Incident response
- Firewall (computing)
- Cyber threat intelligence
Kinds (9)
- Network intrusion detection system (NIDS)
- Host-based intrusion detection system (HIDS)
- Wireless IDPS
- Network behaviour analysis (NBA)
- Intrusion prevention system (IPS)
- Signature-based detection
- Anomaly-based detection
- Stateful protocol analysis
- Reputation-based detection
wired (5)
- Malware
- Security operations center (SOC)
- Security information and event management (SIEM)
- Triage function (single point of contact)
- Collectors, aggregators and agents
Notes published science
- Bodies made the same bargain long before computers: when a miss costs far more than a false alarm, 'optimal systems tend to produce false alarms', which is why fever, pain and panic can look like overreactions (the smoke detector principle, src:nesse-2025-failures). Network defenders choose it too, cutting misses 'at the cost of increasing false positives' (src:nist-sp800-94).
- The arithmetic of rare events sinks most alarms: at a site writing about a million audit records a day with one or two real break-in attempts, even a false alarm rate of 1 in 100,000 means only about two alarms in three are real, and with a more realistic detector barely half (src:axelsson-2000). One SOC analyst puts it bluntly: 'We know 99% of the alarms we generate are false positives, but we still have to look at them' (src:alahmadi-2022-soc).
- A network monitor and the computer it guards can disagree about what arrived. Send a packet the IDS accepts but the host throws away, or the reverse, and the IDS sees a different conversation from the real one; in 1998 every commercial network IDS tested fell to both tricks (src:ptacek-newsham-1998). Snort's author described it as filling an 'ecological niche' (src:roesch-1999-snort).
Layers
The ways it takes in its world, where they meet, and how it acts back, each from L1 (where it meets the world) up.
the boardroom (shared floors)
- L5: Alerting and recording: alerts and anomaly records written, events logged to disk and database servers, correlated across sensors at management servers, sent to consoles, syslog or a SIEM (Bro's policy-script handlers; Snort's logging and alerting subsystem)
- L6: Validation and triage: analysts in a security operations centre judge each alarm (true, false, or a benign trigger), give it context and priority, and close the semantic gap between an anomaly and an actionable report; their sustained attention bounds what is seen
- L7: The security policy and what to do: what counts as an intrusion (violations of security, acceptable-use or standard practice), expressed in rule sets and policy scripts; the decision to respond, block, tune or report
the intrusion detection system as security practitioners describe it: monitored activity up to the security policy it serves (reader's view)
- L1: The monitored activity: network traffic, host events, files, flows, radio frames; intruders and benign users alike
- L2: Sensors and agents: taps and span ports, packet capture, host agents, audit trails, wireless and flow sensors
- L3: Decoding and reduction: packets decoded layer by layer, streams reassembled, reduced to events and audit records
- L4: Detection: signatures, anomaly profiles, stateful protocol analysis, thresholds and lists
- L5: Alerts and records: logged, correlated at management servers, sent to consoles and SIEMs; in-line systems also block
- L6: Validation and triage by analysts: true alarm, false alarm or benign trigger; context; priority
- L7: The security policy and the response: what counts as an intrusion, and what to do about it (respond, block, tune, report)
way in: Critical files and their attributes
- L1: Critical files: system binaries, configuration, ownership and permissions
- L2: Snapshots and message digests taken periodically
- L4: Digests compared with reference values; any difference flagged, after the fact
way in: Host audit records (logins, commands, file and device access)
- L1: Subjects (users, processes) acting on objects (files, commands, devices)
- L2: Audit records generated by the target system for each action performed or attempted
- L3: Records collected daily, consolidated and sorted into sessions
- L4: Each record matched against the subject's profile: counters, means and standard deviations, Markov transitions
way in: Network packets on the monitored link
- L1: Traffic to and from every device on the segment
- L2: Captured at a tap or switch port, copied promiscuously by libpcap and a kernel filter
- L3: Decoded layer by layer and reduced to events (Snort's decoder, Bro's event engine)
- L4: Matched against the rule set (signatures), protocol-state models, or a learned model of normal traffic
way in: New signatures and rule updates
- L1: New exploits revealed (e.g. on the Bugtraq mailing list)
- L2: Rules written in a simple per-packet test language
- L3: Released by the vendor for all its products; updated on time
- L4: Loaded into the detection engine's rule chains, the port every packet is matched against
way in: People noticing (vigilant users, security officers)
- L1: A user behaving strangely; suspicious messages or activity
- L2: Noticed by security officers or vigilant users
way in: Reputation and shared indicators (blacklists, hot lists)
- L1: Hosts, addresses and files already seen in attacks elsewhere
- L2: Written into blacklists (hot lists) and whitelists
- L4: Observed entities matched against the lists; matches blocked or raised in priority
way in: System calls of privileged processes
- L1: Privileged processes running on the host
- L2: A host agent traces which processes call what
- L3: The trace cut into short sequences (windows of six calls)
- L4: Each sequence compared with the normal database; mismatches counted in a locality frame
way in: The monitor's own load (packets received and dropped)
- L1: The monitor's own packet-filter buffer and network interface
- L2: Counts kept of packets received, dropped by the filter, dropped by the interface
- L3: A periodic net stats update event
- L4: Scripts read it to judge whether the monitor is becoming overloaded
way in: Touches on decoys (honeypots)
- L1: Attackers' probes reaching a decoy that looks valuable
- L2: The decoy, isolated and monitored
- L4: Any use is suspect: its value comes from attackers using it
way in: Traffic flows (who talked to whom, how much)
- L1: Communication sessions between hosts
- L2: Flow records (addresses, ports, packet and byte counts, timestamps) from routers or sensors
- L4: Compared with the network's profile against thresholds (web at 13% of border bandwidth)
way in: Wireless frames on radio channels
- L1: Wireless LAN traffic on the radio channels in range
- L2: Sensors hop between channels a few times per second (channel scanning)
- L4: Rogue access points, unauthorised stations and poorly secured devices recognised
way out: Alert to an administrator
- L1: Shown to the administrator
- L3: Carried by syslog, popup messages or the management network to a console
- L5: Alert raised on a match or deviation
way out: Block in-line (drop, reset, throttle)
- L1: Malicious packets dropped, connection reset, offending address blocked
- L3: Inline firewalling, throttling, or TCP resets to both ends
- L5: Prevention action chosen for the alert
way out: Incident response by people
- L1: Telephoning the site the attack comes from; minimising damage
- L7: Decide to trace and contain
way out: Reconfigure firewalls and routers
- L1: Firewalls and routers reconfigure themselves to block certain activity
- L5: Decision to block a source elsewhere
way out: Reports and threat documentation
- L1: Logged threat information used to educate management
- L7: Document the existing threat; find policy problems
way out: Tuning (thresholds, whitelists, profiles, rules)
- L4: Thresholds, black- and whitelists, profiles changed; static profiles regenerated
- L7: Decide to improve accuracy for this environment