{"id":{"repo_id":"mcmaster","oai_identifier":"oai:macsphere.mcmaster.ca:11375/33906"},"canonical_url":"https://search.ndltd.org/etd/mcmaster/oai:macsphere.mcmaster.ca:11375/33906","repository":{"repo_id":"mcmaster","name":"McMaster University","base_url":"https://macsphere.mcmaster.ca/server/oai/request"},"display":{"title":"Hybrid Syscall Modeling for Explainable Container Intrusion Detection","abstract":"Host-based intrusion detectors for containers learn a profile of benign system call behaviour and flag recordings that deviate from it. What they give the analyst is a score. It says that a recording is unusual, but not what about it is unusual. This thesis asks whether a strict partial order learned from the same benign traces can attach a reason to those alarms, in the form of a named ordering constraint that the flagged recording reversed. The detector is left unchanged, and no second opaque model is introduced. A STIDE n-gram detector is paired with a partial-order explanation layer over system call names. The two are trained separately and meet only at reporting. Precedence evidence from benign traces is noisy and contains cycles, so turning it into a strict partial order is an approximation problem with a known hardness result. The thesis applies the randomised R◇ construction of Janicki and Liu to relations derived from program behaviour, and introduces a saturated recovery variant. The pipeline is evaluated on all fifteen scenarios of LID-DS 2021. The explanation layer attaches a violated rule to the detector’s alarms in six scenarios, and which six follows from what the exploit makes the server do. Where it makes the server perform an operation absent from benign traffic, the order of its system calls changes and the layer has something to report. Where it changes only the data a call carries, the order may remain intact and the layer stays silent. The rules are examined for concentration, stability across seeds, and position on the attack timeline. Across the six scenarios, 99.71 percent of the reversed observations fall inside the documented attack window. This establishes temporal alignment. It does not establish that the exploit caused the reversal.","abstract_html":"Host-based intrusion detectors for containers learn a profile of benign system call behaviour and flag recordings that deviate from it. What they give the analyst is a score. It says that a recording is unusual, but not what about it is unusual. This thesis asks whether a strict partial order learned from the same benign traces can attach a reason to those alarms, in the form of a named ordering constraint that the flagged recording reversed. The detector is left unchanged, and no second opaque model is introduced. A STIDE n-gram detector is paired with a partial-order explanation layer over system call names. The two are trained separately and meet only at reporting. Precedence evidence from benign traces is noisy and contains cycles, so turning it into a strict partial order is an approximation problem with a known hardness result. The thesis applies the randomised R◇ construction of Janicki and Liu to relations derived from program behaviour, and introduces a saturated recovery variant. The pipeline is evaluated on all fifteen scenarios of LID-DS 2021. The explanation layer attaches a violated rule to the detector’s alarms in six scenarios, and which six follows from what the exploit makes the server do. Where it makes the server perform an operation absent from benign traffic, the order of its system calls changes and the layer has something to report. Where it changes only the data a call carries, the order may remain intact and the layer stays silent. The rules are examined for concentration, stability across seeds, and position on the attack timeline. Across the six scenarios, 99.71 percent of the reversed observations fall inside the documented attack window. This establishes temporal alignment. It does not establish that the exploit caused the reversal.","abstract_has_math":false,"creators":["Gurbuz, Muhammet Bekir"],"institution":null,"degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":"Computing and Software","school":null,"contributors":[],"advisors":["Janicki, Ryszard","Bandur, Victor"],"committee_chairs":[],"committee_members":[],"year":2026,"date_issued":"2026","date_published":"2026","updated_at":"2026-11-18T19:37:28Z","subjects":["intrusion detection","system calls","containers","anomaly detection","partial orders","explainability","STIDE","LID-DS"],"languages":["en"],"rights":[],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://doi.org/10.71548/25675"],"render_values":[{"text":"https://doi.org/10.71548/25675","href":"https://doi.org/10.71548/25675","code":true}]}],"person_identifier_entries":[]},"links":{"outbound_url":"https://hdl.handle.net/11375/33906","outbound_label":"Handle","outbound_source":"dc:identifier.uri","record_url":"https://hdl.handle.net/11375/33906","record_label":"Repository record","record_kind":"handle","record_source":"dc:identifier.uri","download_url":null,"download_label":null,"download_source":null,"access_status":null,"embargo_until":null,"access_note":null},"source_record":{"url":"https://macsphere.mcmaster.ca/server/oai/request?verb=GetRecord&metadataPrefix=dim&identifier=oai%3Amacsphere.mcmaster.ca%3A11375%2F33906","prefix":"dim"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Janicki, Ryszard","Bandur, Victor"]},{"key":"dc:contributor.department","label":"Department","values":["Computing and Software"]},{"key":"dc:creator","label":"Author","values":["Gurbuz, Muhammet Bekir"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.accessioned","label":"Dc Date Accessioned","values":["2026-10-29T13:03:41Z"]},{"key":"dc:date.issued","label":"Date","values":["2026"]},{"key":"dc:type","label":"Dc Type","values":["Thesis"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["intrusion detection","system calls","containers","anomaly detection","partial orders","explainability","STIDE","LID-DS"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language.iso","label":"Language (ISO)","values":["en"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://hdl.handle.net/11375/33906","https://doi.org/10.71548/25675"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["Host-based intrusion detectors for containers learn a profile of benign system call behaviour and flag recordings that deviate from it. What they give the analyst is a score. It says that a recording is unusual, but not what about it is unusual. This thesis asks whether a strict partial order learned from the same benign traces can attach a reason to those alarms, in the form of a named ordering constraint that the flagged recording reversed. The detector is left unchanged, and no second opaque model is introduced. A STIDE n-gram detector is paired with a partial-order explanation layer over system call names. The two are trained separately and meet only at reporting. Precedence evidence from benign traces is noisy and contains cycles, so turning it into a strict partial order is an approximation problem with a known hardness result. The thesis applies the randomised R◇ construction of Janicki and Liu to relations derived from program behaviour, and introduces a saturated recovery variant. The pipeline is evaluated on all fifteen scenarios of LID-DS 2021. The explanation layer attaches a violated rule to the detector’s alarms in six scenarios, and which six follows from what the exploit makes the server do. Where it makes the server perform an operation absent from benign traffic, the order of its system calls changes and the layer has something to report. Where it changes only the data a call carries, the order may remain intact and the layer stays silent. The rules are examined for concentration, stability across seeds, and position on the attack timeline. Across the six scenarios, 99.71 percent of the reversed observations fall inside the documented attack window. This establishes temporal alignment. It does not establish that the exploit caused the reversal."]},{"key":"dc:description.degree","label":"Dc Description Degree","values":["Master of Science (MSc)"]},{"key":"dc:description.sponsorship","label":"Dc Description Sponsorship","values":["Ministry of National Education of the Republic of Türkiye (YLSY scholarship)"]},{"key":"dc:title","label":"Title","values":["Hybrid Syscall Modeling for Explainable Container Intrusion Detection"]},{"key":"others:access_status","label":"Others Access Status","values":["open.access"]}]}],"canonical_facts":{"dc:contributor.advisor":["Janicki, Ryszard","Bandur, Victor"],"dc:contributor.department":["Computing and Software"],"dc:creator":["Gurbuz, Muhammet Bekir"],"dc:date.accessioned":["2026-10-29T13:03:41Z"],"dc:date.issued":["2026"],"dc:description.abstract":["Host-based intrusion detectors for containers learn a profile of benign system call behaviour and flag recordings that deviate from it. What they give the analyst is a score. It says that a recording is unusual, but not what about it is unusual. This thesis asks whether a strict partial order learned from the same benign traces can attach a reason to those alarms, in the form of a named ordering constraint that the flagged recording reversed. The detector is left unchanged, and no second opaque model is introduced. A STIDE n-gram detector is paired with a partial-order explanation layer over system call names. The two are trained separately and meet only at reporting. Precedence evidence from benign traces is noisy and contains cycles, so turning it into a strict partial order is an approximation problem with a known hardness result. The thesis applies the randomised R◇ construction of Janicki and Liu to relations derived from program behaviour, and introduces a saturated recovery variant. The pipeline is evaluated on all fifteen scenarios of LID-DS 2021. The explanation layer attaches a violated rule to the detector’s alarms in six scenarios, and which six follows from what the exploit makes the server do. Where it makes the server perform an operation absent from benign traffic, the order of its system calls changes and the layer has something to report. Where it changes only the data a call carries, the order may remain intact and the layer stays silent. The rules are examined for concentration, stability across seeds, and position on the attack timeline. Across the six scenarios, 99.71 percent of the reversed observations fall inside the documented attack window. This establishes temporal alignment. It does not establish that the exploit caused the reversal."],"dc:description.degree":["Master of Science (MSc)"],"dc:description.sponsorship":["Ministry of National Education of the Republic of Türkiye (YLSY scholarship)"],"dc:identifier.uri":["https://hdl.handle.net/11375/33906","https://doi.org/10.71548/25675"],"dc:language.iso":["en"],"dc:subject":["intrusion detection","system calls","containers","anomaly detection","partial orders","explainability","STIDE","LID-DS"],"dc:title":["Hybrid Syscall Modeling for Explainable Container Intrusion Detection"],"dc:type":["Thesis"],"others:access_status":["open.access"]},"updated_at":"2026-11-18T19:37:28Z"}