Back to ARUMULLA's profile
10views

Project Aegis: AI Cyber Defense Assistant

An AI-powered SOC teammate that detects network anomalies, clusters threat behaviors, and translates raw telemetry into actionable, plain-English incident summa

ARUMULLA ASHISHPAUL REDDY NITANP
  • Expert reviewed
Project Aegis: AI Cyber Defense Assistant

94%

Threat Detection Accuracy (on UNSW-NB15 testing split)

99%

Reduction in raw alert noise for human analysts

< 2s

AI Incident Summary Generation Time

Overview

The Inspiration: Modern Security Operations Centers (SOCs) are drowning in data. When an attack happens at 2:17 AM, analysts are bombarded with thousands of alerts, most of which are false positives. Aegis was built to shift the paradigm from "human versus hacker" to AI-driven proactive defense, targeting alert fatigue directly.

What it does: Aegis acts as a Tier-1 SOC Analyst that never sleeps. It ingests network traffic and uses a dual-engine machine learning approach to flag malicious behavior. Instead of just throwing a high-severity alert at a human, it passes the flagged telemetry through an LLM to generate a plain-English SITREP (Situation Report), explaining exactly what is happening and why the AI flagged it, complete with actionable next steps.

How it was built: The core threat detection engine uses XGBoost trained on the mandatory UNSW-NB15 dataset to classify specific attack categories (Fuzzers, Backdoors, DoS, etc.). To catch zero-day attacks, I implemented an Isolation Forest model for behavioral anomaly detection. The backend is built in Python, integrating SHAP (SHapley Additive exPlanations) for Explainable AI (XAI). The interactive SOC dashboard was built using Streamlit, providing real-time attack visualization.

What I learned

I learned that in cybersecurity, an AI's accuracy is meaningless without explainability. Early in development, the model successfully caught anomalies but functioned as a "black box," which security teams inherently distrust. By integrating SHAP values into the dashboard, I realized the tradeoff between complex deep learning models and highly interpretable tree-based models like XGBoost. Giving the analyst the exact network feature that triggered the alert (e.g., an abnormal spike in src_bytes) proved far more valuable than a marginal percentage increase in raw accuracy.

AI tools used

GeminiGitHub CopilotChatGPT

Lead AI/ML Engineer and Cybersecurity Developer

Links & files

Artifacts

1