Garuda AI: Intelligent Enterprise Insider Threat Detection
A machine learning-driven security platform designed to identify malicious internal activity, unauthorized privilege escalation, and abnormal data exfiltration using historical behavioral baselines.
1. The Problem
Traditional perimeter firewalls and endpoint antivirus products guard against external intruders, but they remain essentially blind to threats from verified credentials. An authorized employee or an attacker operating with compromised corporate credentials can silently query confidential data stores, copy files to unmonitored endpoints, or escalate privileges without triggering simple rule-based alarms.
Enterprise SOC (Security Operations Center) analysts drown in thousands of daily alerts, creating fatigue and allowing sophisticated insider attacks to go undetected for months.
2. Intended Users & Deployment Scope
SOC Security Analysts
Requires contextualized threat vectors rather than raw log dumps, highlighting deviations with explainable feature importance.
Chief Information Security Officers (CISOs)
Needs a high-level posture dashboard revealing organization-wide risk hotspots and departments with anomalous access trends.
Compliance & Data Audit Teams
Requires auditable traces when sensitive personal records or intellectual property repositories are accessed outside operating hours.
3. Proposed & Implemented Solution
Garuda AI implements User and Entity Behavior Analytics (UEBA). It ingests system audit logs, normalizes heterogeneous log formats, aggregates activities into chronological user session vectors, and feeds them into statistical anomaly algorithms to detect high-entropy deviation from historic baseline behavior.
4. Neel's Role in the Hackathon Team
In our multi-member team at the FinSpark '26 national hackathon:
- Authored Python data ingestion routines using Pandas to clean and parse authentication and access log samples into normalized time-series tables.
- Engineered heuristic behavioral features: login frequency outside business hours, uncommon IP geolocations, and sudden spikes in outbound payload sizes.
- Assisted in benchmarking unsupervised Isolation Forest algorithms against synthetic insider threat datasets (CERT insider threat data subsets).
- Co-presented the prototype to jury evaluation, resulting in our team being honored among the Top 20 national finalist teams.
5. Engineering Challenges & Trade-offs
Extreme Class Imbalance
Legitimate system actions account for >99.9% of all corporate logs. Standard supervised classifiers fail drastically. Switched to unsupervised density estimation and anomaly scoring to isolate true outliers without over-alerting.
Legitimate Work Habits vs Malicious Shifts
Engineers often pull late-night shifts before major releases. Hardcoded time thresholds generated severe false alarms. Implemented rolling 14-day individual baselines to accommodate natural variations.
6. Verifiable Results & Status
Hackathon Milestone Confirmed
Garuda AI was conceived, architected, and prototyped during FinSpark '26, where the project earned our team a prestigious selection into the Top 20 teams nationwide. The source code is maintained on GitHub as an open-source technical prototype.