An Experiment: Network Anomaly Lifecycle

This note is to be removed before publishing as an RFC. Discussion of this document takes place on the Network Management and Operations Area Working Group Working Group mailing list (nmop@ietf.org), which is archived at https://mailarchive.ietf.org/arch/browse/nmop/. Source for this draft and an issue tracker can be found at https://github.com/network-analytics/draft-netana-nmop-network- anomaly-lifecycle.

This document is experimental. The main goal of this document is to propose an iterative lifecycle process to network anomaly detection by proposing a data model for metadata to be addressed at different lifecycle stages. The experiment consists of verifying whether the approach is usable in real use case scenarios to support proper refinement and adjustments of network anomaly detection algorithms. The experiment can be deemed successful if validated at least with an open-source implementation successfully applied with real networks.

The main objective of a network anomaly detection system is to identify Relevant States of the network as those are states that could lead to problems or might be clear indications of problem already happening. In order to formally define the statement above, some definitions from are reported in the following:

Network Anomaly:: An unusual or unexpected event or pattern in network data in the forwarding plane, control plane, or management plane that deviates from the normal, expected behavior.
Network problem:: A state regarded as undesirable and which may require remedial action.
State:: A particular Condition that something (e.g., a Resource) has (i.e., it is in a State) at a specific time.
Relevant State:: A state that have relevancy for network operators, as those are states that could lead to problems or might be clear indications of problem already happening.
Symptom:: An observable Characteristic, State, Event, or Condition considered as an indication of a Problem or potential Problem.

It is still remarkably difficult to gain a full understanding and a complete perspective of "if" and "how" a relevant state is actually an indication of a problem or it is just unexpected, but has no impact on services and end users. Providers of solutions for network anomaly detection should aim at increasing accuracy, by minimizing false positives and false negatives. Moreover, the behaviour of the network naturally changes over time, when more connectivity services are deployed, more customers on-boarded, devices are upgraded or replaced, and therefore it is almost impossible to identify anomaly detection techniquest that can keep working accurately over time, without changing the detection criterias (or methodologies) over time. This opens up to the necessity of further validating notified relevant states to check if a detected symptom is actually impacting connectivity services: this might require different actors (both human and algorithmic) to act during the process and refine their understanding across the network anomaly lifecycle. Finally, once validation has happened, this might lead to refinements to the logic that is used by the detection, so that this process can improve the detection accuracy over time. Performing network anomaly detection is a process that requires a continuous learning and continuous improvement. Relevant states are detected by aggregating and understanding Symptoms, then validated, confirming that Symptoms actually impacted connectivity services impacting and eventually need to be further analyzed by performing postmortem analysis to identify any potential adjustment to improve the detection capability. Each of these steps represents an opportunity to learn and refine the process, and since implementations of these steps might also be provided by different parties and/or products, this document also contributes a formal data model to capture and exchange Symptom information across the lifecycle.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here. This document makes use of the terms defined in . State Problem Event Alarm Symptom The following terms are used as defined in . Metric Intent The following terms are defined in this document. Annotator: Is a human or an algorithm which produces metadata by describing anomalies with Symptoms. False Positive: Is a detected anomaly which has been identified during the postmortem to be not anomalous. False Negative: Is anomalous but has not been identified by the anomaly detection system.

The above definitions of network problem provide the scope for what to be looking for when detecting network anomalies. Concepts like "desirable state" and "required state" are introduced. This poses the attention on a significant problem that network operators have to face: the definition of what is to be considered "desirable" or "undesirable". It is not always easy to detect if a network is operating in an undesired state at a given point in time. To approach this, network operators can rely on different methodologies, more or less deterministic and more or less sensitive: on the one side, the definition of intents (including Service Level Objectives and Service Level Agreements) which approaches the problem top-down; on the other side, the definition of Symptoms, by mean of solutions like SAIN, and , which approaches the problem bottom-up. At the center of these approaches, there are the so-called Symptoms, explaining what is not working as expected in the network, sometimes also providing hints towards issues and their causes. One of the more deterministic approaches is to rely on Symptoms based on measurable service-based KPIs, for example, by using Service Level Indicators, Objectives and Agreements (). This is the case when rules on SLOs and SLIs are manually defined once and the used afterwards for detection at runtime. However, defining SLOs in a "static way" can bring some challenges as well, related to the dynamic nature of networks and services. Alternative methodologies rely on a more "relaxed" approach to detect symptoms and their impact to services as a way to generate analytical data out of operational data. For instance:

SAIN: introduces the definition and exposure of Symptoms as a mechanism for detecting those concerning behaviors in more deterministic ways. Moreover, the concept of "impact score" has been introduced by SAIN, to indicate what is the expected degree of impact that a given Symptom will have on the services relying on the related subservice to which the Symptom is attached.

Daisy: introduces the concept of concern score to indicate what is the degree of concern that a given Symptom could cause a degradation for a connectivity service.

In general, defining boundaries between desirable vs. undesirable in an accurate fashion requires continuous iterations and improvements coming from all the stages of the network anomaly detection lifecycle, by which network engineers can transfer what they learn through the process into new Symptom definitions and, ultimately, into refinements of the detection algorithms.

The lifecycle of a network anomaly can be articulated in three phases, structured as a loop: Detection, Validation, Refinement.

| Detection | ---------+ Adjustments | +-------------+ | Symptoms | | | v +------------+ +------------+ | Refinement |<--------------------- | Validation | +------------+ Problem +------------+ Confirmation ]]> Each of these phases can either be performed by a network expert or an algorithm or complementing each other. The network anomaly metadata is generated by an annotator, which can be either a human expert or an algorithm. The annotator can produce the metadata for a network anomaly, for each stage of the cycle and even multiple versions for the same stage. In each version of the network anomaly metadata, the annotator indicates the list of Symptoms that are part of the network anomaly taken into account. The iterative process is about the identification of the right set of Symptoms.

The Network Anomaly Detection stage is about the continuous monitoring of the network through Network Telemetry and the identification of Symptoms. One of the main requirements that operator have on network anomaly detection systems is the high accuracy. This means having a small number of false negatives, Symptoms causing connectivity service impact are not missed, and false positives, Symptoms that are actually innocuous are not picked up. As the detection stage is becoming more and more automated for production networks, the identified Symptoms might point towards three potential kinds of behaviors: i. those that are surely corresponding to an impact on connectivity services, (e.g. the breach of an SLO), ii. those that will cause problems in the future (e.g. rising trends on a timeseries metric hitting towards saturation), iii. those or which the impact to connectivity services cannot be confirmed (e.g. sudden increase/decrease of timeseries metrics, anomalous amounts of log entries, etc.). The first category requires immediate intervention (a.k.a. the problem is "confirmed"), the second one provides pointers towards early signs of an problem potentially happening in the near future (a.k.a. the problem is "forecasted"), and the third one requires some analysis to confirm if the detected Symptom requires any attention or immediate intervention (a.k.a. the problem is "potential"). As part of the iterative improvement required in this stage, one that is very relevant is the gradual conversion of the third category into one of the first two, which would make the network anomaly detection system more deterministic. The main objective is to reduce uncertainty around the raised alarms by refining the detection algorithms. This can be achieved by either generating new Symptom definitions, adjusting the weights of automated algorithms or other similar approaches.

The key objective for the validation stage is clearly to decide if the detected Symptoms are signaling a real problem (a.k.a. requires action) or if they are to be treated as false positives (a.k.a. suppressing the alarm). For those Symptoms surely having impact on connectivity services, 100% confidence on the fact that a network problem is happening can be assumed. For the other two categories, "forecasted" and "potential", further analysis and validation is required.

After validation of a problem, the service provider performs troubleshooting and resolution of the problem. Although the network might be back in a desired state at this point, network operators can perform detailed postmortem analysis of network problems with the objective to identify useful adjustments to the prevention and detection mechanisms (for instance improving or extending the definition of SLIs and SLOs, refining concern/impact scores, etc.), and improving the accuracy of the validation stage (e.g. automating parts of the validation, implementing automated root cause analysis and automation for remediation actions). In this stage of the lifecycle it is assumed that the problem is under analysis. After the adjustments are performed to the network anomaly detection methods, the cycle starts again, by "replaying" the network anomaly and checking if there is any measurable improvement in the ability to detect problems by using the updated method.

The information that is produced at each stage needs to be persisted and retrieved to perform the network anomaly lifecycle. The lifecycle begins with the detector notifying anomalies to the "Alarm and Problem Management System" and to the "Post-mortem System" according to (see ). In this case the Post-mortem system is identified as the Label Store. Once the notification arrives to the Label Store, the anomaly label is persisted. In the following stages (i.e. validation and refinement), the information about the labels are retrieved, reviewd, modified and persisted again, generating every time a new version of the same annotation, or tagging the annotation as irrelevant, if it would be necessary to remove it. In the following sections, the following are defined: * a state machine for a label * a YANG data model for the notification sent by the Detector to the Label Store * a YANG data model to the define the interrogation (and retrieval) of the labels from the label store.

In the context of this document, from a network anomaly detection point of view a network problem is defined as a collection of interrelated Symptoms, as specified in . The understanding of a network problem can change over time. Moreover, multiple actors are involved in the process of refining this understanding in the different phases. From this perspective, a problem can be refined according to the following states ().

| | Forecasted | | Potential | | | | | +------------+ +-----------+ | | | +--------|--Detection---|-------+ | | | | | +-------+ | +------- ----- + | | Final | | | | +---^---+ | | | | | | | | | v | | | +-----------Validation------------+ | +-----------------------+ | | +-----------+ | | | | | | | | Problem | | Problem | | | | +-----------------+ | | | | Discarded | | Confirmed |<-|---+ | | Detection | | | | +-----|-----+ +-----------+ | | | Adjusted |-------+ +---------------------------------+ | +--------^--------+ | | | | | | | | | | | +---v---+ | | | | | Final | | | | | +-------+ | | +---------|--------+ | | | | Problem | | | | | Analyzed |<-|-----------------------------------+ | +------------------+ | +-------Refinement------+ ]]> The knowledge gained at each stage is codified as a list of anomaly labels that can be stored on a Label Store ( see for a reference).

file "ietf-relevant-state@2025-03-03.yang" module ietf-relevant-state { yang-version 1.1; namespace "urn:ietf:params:xml:ns:yang:ietf-relevant-state"; prefix rsn; import ietf-yang-types { prefix yang; reference "RFC 6021: Common YANG Data Types"; } import ietf-inet-types { prefix inet; reference "RFC 6991: Common YANG Data Types"; } organization "IETF NMOP Working Group"; contact "WG Web: WG List: Authors: Vincenzo Riccobene Thomas Graf Wanting Du Alex Huang Feng "; description "This module defines the relevant-state container and notifications to be used by a network anomaly detection system. The defined objects can be used to augment operational network collected observability data and analytical problem data equally. Describing the relevant-state of observed symptoms. Copyright (c) 2025 IETF Trust and the persons identified as authors of the code. All rights reserved. Redistribution and use in source and binary forms, with or without modification, is permitted pursuant to, and subject to the license terms contained in, the Revised BSD License set forth in Section 4.c of the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info). This version of this YANG module is part of RFC XXXX; see the RFC itself for full legal notices."; revision 2025-03-03 { description "Initial version"; reference "RFC XXX: Semantic Metadata Annotation for Network Anomaly Detection"; } typedef score { type uint8 { range "0 .. 100"; } description "Number that indicates a score between 0 and 100"; } identity network-anomaly-state { description "Base identity for representing the state of the network anomaly"; } identity detection { base network-anomaly-state; description "A problem reached detection state"; } identity validation { base network-anomaly-state; description "A problem reached validation state"; } identity refinement { base network-anomaly-state; description "A problem reached refinement state"; } identity problem-forecasted { base detection; description "A problem has been forecasted, as it is expected that the indicated list of symptoms will impact a service in the near future"; } identity problem-potential { base detection; description "A problem has been detected with a confidence lower than 100%. In order to confirm that this set of symptoms are generating service impact, it requires further validation"; } identity problem-confirmed { base validation; description "After validation, the problem has been confirmed"; } identity discarded { base validation; description "After validation, the network anomaly has been discarded, as there is no evindence that it is causing an problem"; } identity analyzed { base refinement; description "The anomaly detection went through analysis to identify potential ways to further improve the detection process in for future anomalies"; } identity adjusted { base refinement; description "The network anomaly has been solved and analysed. No further action is required."; } identity pattern { description "Pattern identified by the Detector."; } identity drop { description "Drop of the value"; } identity spike { description "Spike of the value"; } identity mean-shift { description "Shift of the mean of the value"; } identity seasonality-shift { description "Shift of the seasonality of the value"; } identity trend { description "Trend exhibited by the value"; } identity other { description "Any other type of pattern"; } grouping relevant-state-grouping { description "Relevant State is a state that could lead to problems or might be clear indications of problem already happening"; leaf description { type string; description "Textual description of the fault"; } leaf start-time { type yang:date-and-time; mandatory true; description "Date and time indicating the beginning of the fault"; } leaf end-time { type yang:date-and-time; description "Date and time indicating the end of the fault"; } leaf concern-score { type score; mandatory true; description "Score indicating the degree of concern in relation to the overall relevant state."; } } grouping annotator-grouping { description "Annotator represents the entity that produced the annotation (it is either a human or an algorithm)"; leaf name { type string; mandatory true; description "Name of the annotator (either user or algorithm) If it is an algorithm, the name can also include the version."; } choice annotator-type { description "An annotator can be either a human user or a programmatic entity, such as an algorithm"; case human { leaf human { type empty; description "This option is used if a human provided the label"; } } case algorithm { leaf algorithm { type empty; description "This option is used if a software provided the label"; } } } } grouping anomaly-grouping { description "List of anomalies that are part of the relevant state"; list anomalies { key "id version"; description "List of Anomaly instances"; leaf id { type yang:uuid; description "Unique ID of the anomaly"; } leaf uri { type inet:uri; description "URI to viusalize the analytical metrics of the anomaly."; } leaf version { type yang:counter32; description "Version of the anomaly metadata object. It allows multiple versions of the metadata to be generated in order to support the definition of multiple problem objects from the same source to facilitate improvements overtime"; } leaf state { type identityref { base network-anomaly-state; } mandatory true; description "State of the anomaly"; } leaf description { type string; description "Textual description of the anomaly"; } leaf start-time { type yang:date-and-time; mandatory true; description "Date and time indicating the beginning of the anomaly A detection system will alwasys set a start time, as it represents the moment in time from which the behaviour of the monitored system is considered to be anomalous with respect its expected behaviour"; } leaf end-time { type yang:date-and-time; description "Date and time indicating the end of the anomaly. This field is indicated as non mandatory, as it could be the case that the anomaly is still happening at the time of generation of the label"; } leaf confidence-score { type score; mandatory true; description "Score indicating how confident was the detector while considering the given anomaly as part of the relevant event"; } leaf pattern { type identityref { base pattern; } description "Pattern describes the type of pattern that was detected by the annotator (e.g. spike, drop, mean-shift, etc.)"; } container annotator { presence "It specifies an annotator for the anomaly"; description "Annotator represents the entity that produced the annotation"; uses annotator-grouping; } container symptom { presence "It specifies the symptom for the anomaly"; description "An observable Characteristic, State, Event, or Condition considered as an indication of a Problem or potential Problem"; leaf id { type yang:uuid; mandatory true; description "Unique ID of the symptom type"; } leaf concern-score { type score; mandatory true; description "Score indicating the degree of concern in relation to the specific symptom. Each symptom will carry a certain degree of concern that is specific to the symptom."; } } container service { presence "It specifies the service (or the monitored entity) affected by the anomaly"; description "Service represents the entity that is monitored, which for instance can be a connectivity servise, such as an L3VPN"; leaf id { type yang:uuid; mandatory true; description "Unique ID of the service (or monitored entity) This is supposed to be augmented by other modules that want to define the service affected by the anomaly"; } } } } notification relevant-state-notification { description "Notification of a relevant state that can be sent by the anomaly detection system to the postmortem management system or to the incident management system"; uses relevant-state-grouping; uses anomaly-grouping; } container relevant-state { description "A Relevant State is a state that have relevancy for network operators, as those are states that could lead to problems or might be clear indications of problem already happening"; leaf id { type yang:uuid; mandatory true; description "Unique ID of the relevant state It is unique in the scope of the Label Store"; } uses relevant-state-grouping; uses anomaly-grouping; } }

]]>



        
          The data model provides support for "human-in-the-loop", allowing 
          for network experts to validate and adjust network anomaly 
          labels and detection systems. An example of human-in-the-loop 
          has been demonstrated with Antagonist , by building a User Interface that 
          interacts with an API based on this data model.
        
        
          The base for the modules is the relevant-state data model.
          Relevant state is at the root of the data model, with its parameters 
          (ID, description, start-time, end-time) and a collection of anomalies.
          This allows the relevant state to be considered as a container of 
          anomalies.
        
        
          Each anomaly is characterized by some intrinsic fields (such as 
          id, version, state, description, start-time, end-time, 
          confidence score and pattern)
          Particularly the confidence score is a measure of how confident 
          was the detector in considering the given anomaly as an anomalous 
          behaviour.
        
        
          Each anomaly also include the symptom and the service container. 
          These containers are placeholders to represent the information 
          about the symptom (what is exaclty happening as anomalous behaviour)
          and the connectivity service (what entity is affected by the anomaly).
          In particular, for what concerns the symptom, a concern score is 
          defined as necessary field, which has the meaning of expressing how 
          much the anomaly is impacting connectivity services.

          In case additional information related to the symptom and to 
          the service need to be provided, augmentation would be the 
          appropriate intended mechanism to do so.

          An example of this is provided in 
          , 
          where an augmentation of both symptom and service is provided 
          for the specific case of anoamly labels related to 
          connectivity services.
        
        
        
        Also a list of various actors that can be involved in the process 
        is presented as following:
          
            In the detection stage:
              
                the detectors can be Network Engineers 
                and/or Automatic detectors 
                (including Rule-based detectors and ML-based detectors)
              
          

          
            In the validation stage:
            
              the validators can be Network Engineers 
              manually validating the labels
            
          

          
            In the refinement stage:
            
              the refiners can be Data Scientists and/or Automatic 
              Refiners (including systems that automatically refine 
              the detection systems, based on the validated labels).
            
          
        
        
        
        The data model that has been defined is used in two YANG modules:
        the relevant-state-notification and the ietf-relevant-state: 
        the notification is primarily used by the Network Anomaly Detector, 
        to notify the "Alarm and Problem Management System" 
        and the "Post-mortem System" 
        (see ); 
        the container instead is used inside the Post-mortem system 
        to exhance anomaly  detection lables between the anomaly 
        detection stages defined above (validation, refinement, detection).



    
      
        This section provides pointers to existing open source
        implementations of this draft. Note to the RFC-editor: Please remove
        this before publishing.
      

      
        An open-source implementation for this draft is called AnTagOnIst
        (Anomaly Tagging On hIstorical data), and it has been implemented in
        order to validate the application of the YANG model defined in this
        draft. Antagonist provides visual support for two important use cases
        in the scope of this document: 
            the generation of a ground truth in relation to symptoms and
            problems in timeseries data

            the visual validation of results produced by automated network
            anomaly detection tools.
          
 The open-source code can be found here: 

        
        As part of the experiment that was conducted with AnTagOnIst, 
        Some main Use Case scenarios have been validated so far:
        
        
          
            Exposure of a GUI for human validation of the labels.
          
          
            Integration with Rule Based anomaly detection systems. 
            In particular the integration with SAIN and 
            Cosmos Bright Lights is ongoing.
          
          
            Integration with ML-based detection systems.
          
        
        
      
    

    
      The security considerations will have to be updated according to
      "https://wiki.ietf.org/group/ops/yang-security-guidelines".
    

    
      
        The authors would like to thank Antonio Roberto for his contribution 
        to the ideas in this draft and Mohamed Boucadair for his review 
        and valuable comments.