AI-BASED SOC ASSISTANT FOR EVENT CORRELATION, LOG ANALYSIS, AND THREAT DETECTION
Abstract
Security Operations Centers (SOCs) process large volumes of endpoint, network, identity, and application telemetry. This paper presents an AI-Based SOC Assistant that combines telemetry ingestion and normalization, temporal event correlation, deterministic threat detection, incident management, AI-assisted investigation, and analyst-facing presentation. The project architecture identifies Splunk Boss of the SOC (BOTS) v1 CSV data (15.8 GB and approximately 13.4 million rows) and a sample JSONL collection (1,508 events) as input sources; these are source-inventory figures and do not establish that the entire BOTS dataset was processed in a benchmark run. Events are parsed, validated using a Pedantic schema, and stored in SQLite. Four correlation strategies group activity by source IP, user, multiple destination ports, and authentication sequence, using configured sliding windows described in the technical dossier as 300–600 seconds. A deterministic engine specifies seven MITRE ATT&CK-mapped detection categories: brute force (T1110), suspicious successful logon (T1078), privilege escalation (T1068), suspicious process execution (T1059.001), port scanning (T1046), lateral movement (T1021), and unauthorized service installation (T1543.003). Related detections are grouped into incidents with severity and lifecycle tracking. A local Ollama model (Qwen 2.5 7B or Llama 3) supports evidence-based summaries and remediation guidance, with an expert fallback described in the project specification. The Stream lit dashboard presents Command Center, AI Workbench, and Telemetry Explorer views through a Fast API service. The technical dossier reports performance measurements, but they are reproduced here as reported values rather than independently verified results. Consequential response actions require authorization, safeguards, and analyst approval.