Obrina Briliyant
Researcher
BSSN
Obrina Briliyant is currently active as a full-time research student in the School of Computer Science and Informatics, Cardiff University, UK. He is also a research associate and an experienced IT auditor in Indonesia, with focus on IT security assessment and audit based on COBIT, ISO 27001, CIS, and OWASP. He is still active in the Indonesian Cyber Security Agency (BSSN) and ISACA Indonesia Chapter. His main interests are automated cyber defense and privacy-preserving systems for IoT environment.
Speaker sessions
Briefings and Arsenals | Confirmed Speakers | No Session Assigned
Placeholder session for confirmed speakers without an assigned session, allowing them to be pushed to Swapcard and Visit.
- 07:30
- Tue
- 01 Dec
Stage:
Briefing Stage 2
Sessions Type:
Keynote
SACA: Responsible AI for IoT Network Security Audit
AI-powered intrusion detection systems have achieved impressive accuracy, but accuracy is not the same as trustworthiness. Most detection systems lack explainability, traceability, and human intervention. The consequence is a high-F1 pipeline that an analyst cannot audit. When a system flags a flow as malicious, there is no mechanism to interrogate why, to verify whether it is grounded in actual packet evidence, or to detect whether the explanation itself has been manipulated.
This session presents SACA (Security Audit Compliance Agent), a prototype that addresses this gap by treating faithfulness measurement and evidence traceability as first-class requirements rather than extras. SACA uses GNN4ID (Farrukh et al., 2025) as its detection backbone, which achieves an F1 score exceeding 0.97 on the CIC-IoT2023 dataset. Malicious flows are then routed to an explanation layer where a human auditor can query the packet data in natural language, with every LLM response grounded in a knowledge graph and scored for faithfulness against the retrieved evidence using the Retrieval-Augmented Generation Assessment Suite (RAGAS).
The session also demonstrates an emerging attack surface: once a generative model is attached to a detection system to narrate verdicts, that explanation layer becomes a target for crafted payloads, manipulated log content, and prompt injection smuggled via packet data. SACA treats this as an adversarial problem and proposes faithfulness scoring as a detection mechanism for manipulation. We demonstrate a live case where the model narrative and the underlying evidence diverge, and the faithfulness score surfaces the discrepancy before the finding reaches a report.
- 16:20
- Tue
- 01 Dec
Stage:
Arsenal 4
Sessions Type:
Keynote