Back to Ayushman's profile
7views

CyberGuard

An AI-powered cybersecurity platform combining threat detection, graph intelligence, and reasoning-based analysis to identify and explain digital threats.

Ayushman Patra
  • Problem, approach and result explained
  • Expert reviewed· 6 reviews
CyberGuard

80%

Multi-layer Threat detection architecture

98.5%

Graph-based Threat intelligence

100%

Explainable AI Contextual threat assessment

Overview

The problem

Today, cyberattacks are no longer limited to suspicious emails or easily identifiable fake websites. Attackers are using AI to create convincing phishing websites, impersonate trusted individuals, generate fraudulent messages, manipulate images and audio, and exploit people's trust in digital platforms.

One major weakness in traditional security systems is that they often analyze threats individually. A domain might have a good reputation, but one of its subdomains could be compromised. A message might look legitimate on its own, but its sender, URL, and associated digital identity could reveal a coordinated scam.

We realized that cybersecurity needs to move beyond simply asking whether something is malicious or safe. It needs to understand the relationships between digital entities, identify suspicious behavior, and explain why something might be dangerous.

That is the problem we are trying to solve with CyberGuard.

The approach

Our solution is CyberGuard, an AI-powered cybersecurity and threat intelligence platform developed under our project Cybernates.

Instead of depending on a single detection model, we are designing a hybrid architecture that combines multiple security techniques.

There are five major components to our approach:

  1. Multi-layer threat detection: Analyzing URLs, domains, websites, SMS, suspicious messages, and manipulated digital media.
  2. Lexical and behavioral analysis: Examining URL structures, suspicious keywords, domain characteristics, redirects, and behavioral indicators.
  3. Graph intelligence: Connecting entities such as domains, subdomains, IP addresses, URLs, senders, and other indicators to identify hidden relationships and suspicious activity.
  4. Reasoning-based AI: Adding contextual analysis to interpret multiple signals and produce understandable threat explanations rather than relying only on a binary safe-or-malicious prediction.
  5. Actionable protection: Presenting threat scores, risk indicators, explanations, and recommended security actions. One important design decision is that we do not want to automatically trust an entire domain just because its parent domain has a good reputation. We want to analyze individual subdomains and their behavior independently.

Our longer-term objective is to extend CyberGuard into a single-user protection system through local firewall and network-monitoring integration.

The result

Our goal is not just to build another cybersecurity dashboard. We want to demonstrate how different detection techniques can work together as one intelligent security system.

Our prototype development is focused on running and validating individual detection components locally, with a unified interface planned for presenting their results.

For our demonstration, we want to show practical scenarios rather than just slides.

For example:

  • A suspicious phishing URL is analyzed and its risk indicators are displayed.
  • A fraudulent SMS is examined for suspicious language, links, and social-engineering patterns.
  • A trusted-looking domain with a suspicious subdomain is investigated separately.
  • Related indicators are connected through graph intelligence.
  • The system provides an explanation of its assessment and suggests appropriate security actions. The important outcome is an architecture designed to combine detection, contextual intelligence, and explainability into a single workflow.

Reflection

Building CyberGuard has taught us that cybersecurity is much more complicated than simply training a machine learning model to classify something as safe or malicious.

One of our biggest challenges is dealing with false positives and false negatives. A newly created website is not necessarily malicious, and an old, trusted domain is not necessarily safe forever.

We also learned that combining multiple AI models does not automatically create a better security system. Each component needs to be tested, validated, and integrated carefully.

Another important challenge is real-time protection. Detecting a threat in an isolated demonstration is different from protecting a user's actual device and network. This is why we are exploring local firewall integration and a single-user deployment architecture.

Going forward, our priorities are improving detection accuracy, validating every module against realistic datasets, strengthening graph-based threat correlation, and building a reliable end-to-end protection workflow.

Our larger vision is to make cybersecurity more proactive, explainable, and accessible, so that users do not have to be cybersecurity experts to understand and protect themselves against digital threats.

Submitted to CYBERGUARD: Trust Nothing, Explain Everything

Your idea in one line

CyberGuard is an autonomous AI-powered Security Operations Center (SOC) and adaptive firewall platform designed for state educational grids, e-governance portals, and enterprise datacenters to detect and neutralize synchronized cyber attacks across both technical networks and multimodal human deception vectors (voice clones, deepfakes, and DLT smishing).

What will be working by the finale?

Autonomous Adaptive Firewall & Netlink Verification: Real-time synthesis of ephemeral Linux nftables drop rules in under 11ms for high-risk IPs ($\ge 85$), verified via kernel Netlink sockets (APPLIED_VERIFIED) with automatic TTL quarantine expiration. Unified SOC Command Center (React 19 + TypeScript): Real-time fleet health scoring (92/100 nominal posture), live threat streams, incident investigation (INC-1042), and sub-millisecond telemetry synchronization via WebSockets. Multimodal Threat Lab: Active detectors for: Voice Clone Audio Deepfakes: 128-band Mel-Spectrogram analysis detecting neural vocoder high-frequency drop-offs ($>8.1\text{ kHz}$) and flat pitch contours. Facial Seam & Video Forensics: 2D Discrete Cosine Transform (DCT) and spatial landmark warping detection. QR Quishing & Telecom Smishing: Decoding high-entropy QR matrices and validating SMS sender headers against TRAI DLT registers with RoBERTa urgency scoring. Ensemble ML Flow Detection Engine: Production XGBoost classifier ($D=42$ flow metrics, 98.4% accuracy), KNN similarity matching across 16,682 attack archetypes, and unsupervised Deep Autoencoder reconstruction for zero-day anomalies. Interactive Attack Simulator Sandbox: Live demonstration of 6 distinct attack scenarios (Port Scan, Phishing masq, Credential Stuffing / Impossible Travel, Volumetric SYN Flood, Audio Voice Clone, and Quishing Blitz) demonstrating machine-speed autonomous mitigation. Grounded AI Security Analyst (Qwen 2.5): Natural language SOC Copilot providing deterministic triage summaries and automated MITRE ATT&CK kill-chain mappings.

Tech stack and hardware you plan to use

Frontend: React 19, TypeScript, Vite 8, TailwindCSS / Calm Enterprise Security Intelligence design system, Lucide React, native WebSockets. Backend & API: Python 3.11/3.13, FastAPI (ASGI), Uvicorn, Pydantic v2, Argon2id cryptography, PyJWT stateless authentication. Kernel & Firewall Infrastructure: Linux nftables (table inet cyberguard_inet), Linux Netlink Sockets (AF_NETLINK), Redis for session invalidation. AI & Machine Learning Models: XGBoost (v4.19): Tabular flow classification trained on UNSW-NB15 & CIC-DDoS2019. PyTorch Deep Autoencoder: Bottleneck reconstruction neural network for zero-day anomalies. $k$-Nearest Neighbors ($k=5$): Case-based reasoning across 16,682 attack archetypes. RoBERTa (Transformers): NLP deceptive urgency & zero-width homoglyph detector. RawNet2 & 128-band Mel-Spectrograms: Audio deepfake & neural vocoder artifact detection. ResNet50 / ViT & 2D DCT: Computer vision facial seam and tampering detection. Heterogeneous RGAT GNN (PyTorch Geometric): Multi-hop infrastructure graph correlation. Qwen 2.5 (3B/7B Instruct): Grounded local AI SOC incident analyst. Hardware & Edge Requirements: Standard x86_64 Linux edge server / gateway node with Linux Kernel 5.4+ (support for nftables & Netlink). 1x NVIDIA GPU (RTX 3060 / T4 or higher) or multi-core CPU for edge neural inference. Network TAP / SPAN mirror port or Suricata/Zeek IPFIX/NetFlow sensor feed.

A diagram, sketch or mock-up of your idea
A diagram, sketch or mock-up of your idea: CyberGuard
What have you built or tested so far?

We have already built a fully functional end-to-end working system and verified it with empirical benchmarks: Complete 17-Screen SOC Frontend: Implemented and operational in React 19 + TypeScript, covering Overview, Live Threat Feeds, Adaptive Firewall, Identity Security, Phishing Sandbox, Multimodal Threat Lab, and GNN Attack Graph. Trained & Serialized ML Models: Primary XGBoost flow model evaluated at 98.4% Accuracy and 97.9% F1-Score on UNSW-NB15/CIC-DDoS with $<2.1\text{ ms}$ inference time. Deep Autoencoder trained on benign flow manifolds flagging zero-days via reconstruction MSE ($\tau = 0.0215$). $k$-NN engine indexed over 16,682 historical attack archetypes. FastAPI Backend Core: All REST endpoints under /api/v1 and WebSocket live telemetry streaming tested and functional with Argon2id RBAC auth. Deterministic Policy & Netlink Pipeline: Verified autonomous policy enforcement loop executing ephemeral firewall drop rules in 11 milliseconds with verified netlink audit receipts. Interactive Attack Simulator: Tested 6 synthetic attack injection scenarios (port scan, DLT smishing spoof, CEO audio voice clone, impossible travel, and DDoS) with live policy mitigation feedback

Pitch deck (PDF)
Pitch deck (PDF)PDF
What's working now?

We are developing **CyberGuard**, an AI-powered cybersecurity platform under Cybernates that aims to detect, analyze, and prevent modern cyber threats through a hybrid intelligence approach. The prototype focuses on multiple threat categories, including phishing websites, malicious URLs and domains, fraudulent messages and SMS, suspicious digital identities, and AI-generated or manipulated media. Our planned architecture combines machine learning-based detection, lexical and behavioral analysis, graph intelligence for connecting suspicious entities and identifying hidden relationships, and reasoning-based AI for contextual threat assessment. We are prioritizing a working single-user demonstration with real-time threat analysis, explainable detection results, and actionable security recommendations. The current focus is integrating and validating individual detection modules locally before bringing them together into a unified CyberGuard experience.

Deck or report (PDF)
Deck or report (PDF)PDF
A photo or screenshot
A photo or screenshot: CyberGuard

AI tools used

ChatGPT

We used ChatGPT and other AI-assisted development tools for technical research, system architecture planning, and implementation support throughout CyberGuard's development. These tools helped us design the hybrid AI-based detection architecture, explore machine learning approaches for phishing and fraud detection, understand graph intelligence and entity relationships, identify reusable open-source cybersecurity repositories, prepare technical documentation (PRD, DRD), troubleshoot code, and plan integration between backend detection modules and the frontend. AI tools were used to accelerate research and development, while the goal remains to validate every implemented feature through actual testing and demonstrations.

Links & files

Artifacts

1

Gallery

3