Research & Technical Details

Patent-pending technology for real-time network threat detection using computer vision on edge hardware.

The Problem

Traditional network intrusion detection systems (IDS) rely on signature matching — comparing packets against known attack patterns. This approach fails against novel attacks, zero-day exploits, and AI-generated attack chains. As autonomous AI systems like Anthropic's Claude Mythos Preview demonstrate the ability to discover and exploit vulnerabilities across every major operating system, signature-based detection becomes fundamentally insufficient.

Our Approach: Network Activity Images

NetworkVision converts raw network traffic into 2D images called Network Activity Images (NAIs), then applies YOLO-style object detection to identify and localize threats in real-time.

NAI Construction

Each NAI is a multi-channel image where:

ChannelDataPurpose
0Packet SizePayload volume per flow
1Packet CountActivity density
2ProtocolTCP/UDP/ICMP distribution
3TCP FlagsConnection state patterns
4DirectionInbound vs outbound
5Inter-Arrival TimeTiming patterns
6Byte RatioAsymmetric transfer detection
7Unique Dest PortsScan behavior
8SYN/ACK RatioConnection anomalies
9Connection DurationPersistent vs ephemeral flows

Rows represent flows (mapped via consistent hashing), columns represent time bins within a 30-second window. The result is a spatial representation where attack patterns form visually distinct signatures.

Detection Architecture

Two-Stage Pipeline

Stage 1 — Binary Classifier: Fast triage model determines if a NAI contains any threat. 83KB, 0.28ms inference on Coral Edge TPU.

Stage 2 — YOLO Localizer: If a threat is detected, the localizer identifies threat type and location within the NAI using bounding-box regression. 376KB, 0.21ms on Coral Edge TPU.

Combined inference: 0.49ms — enabling real-time detection at line rate on a $20 AI chip.

Threat Classes

ClassNAI PatternDescription
Port ScanHorizontal spread across many destination portsReconnaissance activity probing for open services
DoS/DDoSDense vertical bands from many sourcesVolumetric or application-layer denial of service
ExploitConcentrated bursts with unusual flag patternsActive exploitation attempts
Web AttackHTTP-port focused with asymmetric byte ratiosSQL injection, XSS, path traversal
Botnet/C2Periodic beacon patterns to external IPsCommand and control communication
Brute ForceRepeated connections to auth portsCredential stuffing and password attacks
BenignNormal traffic distributionLegitimate network activity

Auto-Labeling Pipeline

NetworkVision includes a proprietary auto-labeling system that generates training data continuously from live network traffic. Traditional signature-based IDS alerts are correlated with behavioral observations to produce labeled datasets — enabling a self-improving detection loop that adapts to new threat patterns over time.

This approach eliminates the need for manual labeling and allows the model to generalize beyond known signatures to detect novel attacks.

Edge Hardware Performance

PlatformBinary ClassifierYOLO LocalizerCombinedPower
Pi 5 CPU0.32ms0.70ms1.02ms~5W
Coral Edge TPU0.28ms0.21ms0.49ms~0.5W
Hailo-10H20 TOPS INT8 — multi-model capable~2.5W

Why Behavioral Detection Matters

Post-Mythos Threat Landscape

AI systems can now autonomously discover and exploit zero-day vulnerabilities. Every novel exploit generates unique byte-level signatures that evade traditional IDS. But the network behavior remains detectable:

• Reconnaissance still produces scan patterns
• Exploitation still requires packet delivery
• Lateral movement creates new internal flows
• C2 channels produce periodic beacons
• Exfiltration creates asymmetric byte ratios

NetworkVision detects these behavioral patterns regardless of the specific exploit, making it resilient to novel and AI-generated attacks.

Status

Patent Pending Provisional Filed April 2026 14 Claims Micro Entity

For partnership inquiries: Partners@NetworkVision.org