Monitoring Systems

• Stephan Edel
• Anne Walther
• presentation date: 01/04/2005

Motivation

• Protective measures like firewalle reduce the attack risk, but cannot totally eliminate it
  • e.g. bogus packer headers can lead to an attack not being detected
• are principally "burglar alarms"
• are employed for automated attack detection -> by monitoring accordance with provided security guidelines / rules
• possible statements produced by an analysis are e.g.:
  • accordance with security guidelines
  • violation of security guidelines
  • suspective behaviour
• about attackers:
  • from the outside
  • from the inside: according to a study by the FBI, the larger part of network attacks originates from its inside (Insiders know more about a systems architecture, protective measures etc.)
  • types: vandalism / espionage / just for fun / find security flaws in cooperation with the host
• maybe a better name would be "`Signature Compare System"' or "`Signature Compare Automaton"'

timing behaviour

• Realtime IDS - reports suspected intrusions which are proceeding (e.g. by network packet analysis)
• Standard IDS - reports sucessful attacks (e.g. by logfile analysis)

architectures

HIDS - host-based IDS

• receive their audit files from a host
• use e.g. log files
• system manipulations can be detected using checksums
• by these mechanisms, attacks can only be detected when they already have occured -> forensic use
• examples:
  • sXid - monitors SUID/GUID files
  • Tripwire - monitors file integrity (employing a database and hash functions to detect changes)

logcheck - logfile analysis by provided strings which are classified as security relevant: denied / failed; security violations: login root refused

NIDS - network-based IDS

• analyzes network data
• NIDS come in different flavours:
  • sensor-based system: has a network interface to "`sniff"' on the traffic
  • agent-based system: agents are installed on the monitored systems to catch the packets (from the different of the TCP/IP stack) and send them to a centralised unit for analyis.
• Position:
  • between "`outside world"' and firewall, but before the internal network -> interesting to get an overview, but not a good choice on the long run because too much data is collected (..every portscan is recorded)
  • behind firewall, inside the local network -> can be used as a controlling device, unusal activities from local hosts are also reported
  • between 2 firewalls -> does not have to react on every single portscan, can be used to monitor inbound as well as outbound traffic
  • multiple sensors -> are useful in larger networks, collected data can be sent to a central analysis unit

hybrid forms

• combinations of HIDS and NIDS
• by partial data analyis on each host the central analysis unit can be relieved of some load

analysis techniques

misuse detection

• works similar to virus scanners
• detection by known signatures
• pattern matching -> incoming pattern is compared to known attack patterns from a database
• signs of an attack:
  • malformed network packets (big / small / characteristic sub-strings in packet data)
  • unusual protocols
  • access to certain ports
• Disadvantage: only known attack methods can be detected. An up-to-date attack database is essential!

State Transition

• another misuse detection technique
• every state represents a state of the system / each transition is an action
• state-transition diagram for every attack pattern
• when system activity occurs, the analysis unit can change the state of the state automaton
• (+) probability and proceeding of an attack can be modelled / some prediction can be made on the attackers next actions
• (+) an attack is modelled on a very abstract basis / attack variants can be modelled with small additional automatons
• (+) coordinated and slowed attacks can be detected!
• next stage of development: colored Petri nets

Anomaly Detection

• is not based on a static dataset
• assumption: every behaviour that is not normal is an attack / suspective behaviour
• the system has to "learn" for severeal weeks what normal behaviour means
• this must be done while the systems integrity is certainz
• an IDS may register e.g.
  • protocols
  • ports
  • attributes of system critical files (rights / users / size / modification date etc.)
• when the learing phase is completed, learned behaviour is considered normal, everything else as anomalous.

Quantitative Analysis

• Treshold Detection:
  • system and user activities are represented by counters:
  • exceeding a treshold is reported as possible attack
  • simple example: system log in, threshold of 3 -> 3 failed logins - user account is blocked
• extension: heuristical treshold analyis
  • treshold is not fixed, but adjusted dynamically
  • in our example: the treshold is computed from the users past login activities
• integrity checking
  • changes to system objects that may not be changed (data, programs, hardware) are monitored

Statistical Analysis

• IDS keeps statistical profiles on each users normal behaviour
• security alert is defined as strong deviation from this normal behaviour
• (+) attackers using a "hijacked" account can be detected
• (+) previously unknown patterns and methods of attack can be recognized
• (-) bad real-time behaviour: all actions must have already happend to see if they are normal
• (-) may pose too many restrictions on users
• (-) may simplify "social monitoring" ("`..user X sends emails after lunch break"')

rule-based systems

• operation similar to statistical systems (rules are defined instead of statistics)
• rules may be user-defined or derived from data analyis
• special subclass: Time-Based-Indictive IDS
  • monitors sequence of system operations
  • an anomaly is detected if order of events is not correct

neuronal networks

• network is trained on "clean" data
• learning ability makes neuronal networks wells suited for anomaly detection
• (-) does not report type of attack
• solution: one network for each type of attack (e.g. SYN-flooding)

Strict Anomaly Detection (Burglar Alarm / Passive Traps)

• assumption: everything that is not explicitely permitted is an attack
• employs pattern matching
• normal system behaviour provided in a database
• every deviant system behaviour is reported as an attack
• (+) fewer patterns must be saved in the database
• (+) updates to the database are only needed when the system is changed