Zero-Trust Network Architecture: Adversarial Attack Detection in ML-Based ZT

System Overview

This document describes the complete architecture of the Zero-Trust Network Adversarial Detection System.

High-Level Architecture

┌─────────────────────────────────────────────────────────────┐
│                 ZERO-TRUST NETWORK SYSTEM                   │
└─────────────────────────────────────────────────────────────┘

┌──────────────────┐
│  Network Flow    │ ← NSL-KDD Dataset (41 features)
│  (NSL-KDD)       │   - Duration, protocol, service, flags
└────────┬─────────┘   - Byte counts, packet stats
         │             - Connection patterns
         ▼
┌──────────────────────┐
│ Feature Extraction   │ ← Preprocessing:
│  & Normalization     │   - Label encoding (categorical)
└────────┬─────────────┘   - StandardScaler (numerical)
         │
         ▼
┌──────────────────────────┐
│ ML Risk Classifier       │ ← Neural Network:
│  (NetworkRiskModel)      │   Input(41) → 128 → 64 → 32 → 1
│                          │   Dropout(0.3), ReLU, Sigmoid
└────────┬─────────────────┘
         │
         ├─→ Risk Score (0-1)
         │
         ▼
┌──────────────────────────────┐
│ Context Enrichment           │ ← Logic-Driven Trust (Posture):
│  - User identity             │   - Patch compliance (0-1)
│  - Device trust (0-1)        │   - Behavioral anomaly (0-1)
│  - Geo risk (0-1)            │   - Verification age (days)
│  - Time-of-day risk          │   - Resource-level sensitivity
│  - Requested segment         │     (Micro-segmentation)
└────────┬─────────────────────┘
         │
         ▼
┌──────────────────────────────────────┐
│ Zero-Trust Policy Engine             │ ← Multi-Factor & Segment Rules:
│                                      │
│  Rules:                              │
│  • Resource ML Threshold → DENY      │ (Micro-segmentation Rule 0)
│  • Global ML risk > 0.8 → DENY       │
│  • Device trust < 0.5 → MFA/DENY     │
│  • Geo risk > 0.7 → MFA              │
│  • Admin segment → strict checks     │
│  • Otherwise → ALLOW/RATE_LIMIT      │
└────────┬─────────────────────────────┘
         │
         ├─→ Decision: ALLOW/DENY/MFA/RATE_LIMIT
         │
         ▼
┌──────────────────────────┐
│ SOC Telemetry Logging    │ ← Audit Trail:
│  - Flow ID               │   - All decisions
│  - User, IP, segment     │   - Risk scores
│  - All scores & decision │   - Context data
│  - Timestamp & reason    │   - JSON export
└──────────────────────────┘

═══════════════════════════════════════════════
        ADVERSARIAL ATTACK LAYER
═══════════════════════════════════════════════

┌──────────────────────────────┐
│ Adversarial Attacker         │ ← Attack Methods:
│  - FGSM on network flows     │   - Gradient-based
│  - PGD multi-iteration       │   - Network constraints
│  - Evasion scenarios         │   - Realistic bounds
└────────┬─────────────────────┘
         │
         ▼
┌──────────────────────────────┐
│ Evasion Evaluation           │ ← Metrics:
│  - Attack Success Rate       │   - Evasion success
│  - Risk score reduction      │   - L2/Linf distance
│  - Policy bypass rate        │   - Detection rates
└──────────────────────────────┘

Component Details

1. Network Data Layer

Purpose: Load and preprocess NSL-KDD network traffic data

Key Features:

41 network flow features
Label encoding for categorical data
StandardScaler normalization
Train/test split handling

Input: Raw NSL-KDD CSV files Output: Preprocessed numpy arrays (X, y)

2. ML Risk Classifier

Purpose: Detect malicious network traffic

Architecture:

Layer 1: Linear(41 → 128) + ReLU + Dropout(0.3)
Layer 2: Linear(128 → 64) + ReLU + Dropout(0.3)
Layer 3: Linear(64 → 32) + ReLU
Layer 4: Linear(32 → 1) + Sigmoid

Performance:

Accuracy: 78.5%
Precision: 97.2%
Recall: 64.1%
F1: 77.3%

3. Context Enrichment

Purpose: Add Zero-Trust metadata to network flows

Context Fields:

user_identity: Simulated user ID
device_trust_score: 0-1 (Derives from posture logic)
device_posture: Raw telemetry (Compliance, Anomaly, Age)
geo_risk_score: 0-1 (geographic risk)
time_of_day_risk: 0-1 (temporal risk)
requested_segment: Resource level (e.g., 'db', 'web')

Trust Sources (Simulated Logic):

Patch Compliance: weighted 40%
Behavioral Anomaly: weighted 40% (Inverse: 1.0 - score)
Verification Age: weighted 20% (Linear decay over 365 days)

4. Zero-Trust Policy Engine

Purpose: Make access control decisions

Policy Rules:

Priority	Condition	Decision	Reason
0	ML Risk > Resource Threshold	DENY	Micro-segmentation breach
1	Global ML Risk > 0.8	DENY	High risk score
2	Device Trust < 0.5 AND ML Risk > 0.4	DENY	Untrusted device + risk
3	Device Trust < 0.5	STEP_UP_AUTH	Require MFA
4	Geo Risk > 0.7	STEP_UP_AUTH	High geo risk
5	Sensitive Segment AND ML Risk > 0.6	STEP_UP_AUTH	Sensitive access
6	Sensitive Segment AND Device Trust < 0.7	DENY	Insufficient trust
7	ML Risk > 0.6	RATE_LIMIT	Moderate risk
8	Default	ALLOW	All checks passed

Resource Thresholds (example):

admin: 0.4
database: 0.5
api: 0.6
internal: 0.7
web: 0.8

Access Decisions:

ALLOW: Full access
DENY: Block access
STEP_UP_AUTH: Require MFA
RATE_LIMIT: Throttle access
ISOLATE: Quarantine

5. Adversarial Attack Layer

Purpose: Test system robustness against evasion

Attack Methods:

FGSM (Fast Gradient Sign Method)

perturbation = epsilon * gradient.sign()
x_adv = x - perturbation  # Minimize loss

PGD (Projected Gradient Descent)

for i in range(num_iter):
    perturbation = alpha * gradient.sign()
    x_adv = x - perturbation
    x_adv = project_to_epsilon_ball(x_adv)

Network Constraints:

Gradient Masking: Categorical features protected from drift
Feature Clipping: Per-feature valid ranges (duration >= 0, etc.)
Integer Rounding: Packet/Byte counts rounded
Protocol Bounds: Statistical range enforcement

Evasion Scenarios:

Slow rate limiting
Port hopping
Mimicry attacks
Packet fragmentation

Data Flow

Normal Flow Processing

1. Network packet arrives
2. Extract 41 features
3. Preprocess (encode, scale)
4. ML risk scoring → 0.7
5. Build context:
   - User: user_abc123
   - Device trust: 0.8
   - Geo risk: 0.2
   - Segment: web
6. Policy evaluation:
   - ML risk 0.7 > 0.6 → RATE_LIMIT
7. Log decision
8. Return: RATE_LIMIT

Adversarial Flow Processing

1. Malicious flow detected (risk: 0.9)
2. Attacker generates adversarial example
3. FGSM perturbation applied
4. Network constraints enforced
5. Adversarial flow submitted (risk: 0.5)
6. Context enrichment:
   - Device trust: 0.4 (low)
   - Geo risk: 0.8 (high)
7. Policy catches via context:
   - Device trust < 0.5 → DENY
8. Attack fails despite ML evasion

Threat Model

Attacker Capabilities

Knowledge:

Black-box access to ML model
Can query model for predictions
Knows feature space

Constraints:

Must maintain valid network flows
Limited perturbation budget
Cannot modify identity/device context

Goals:

Evade ML detection
Bypass Zero-Trust policies
Gain unauthorized access

Defense Mechanisms

Layer 1: ML Detection

High precision (97.2%)
Detects known attack patterns

Layer 2: Zero-Trust Context

Device trust verification
Geo-risk assessment
Segment-based policies

Layer 3: Policy Enforcement

Multi-factor decision making
Defense-in-depth
Adaptive thresholds

System Metrics

Performance Metrics

Throughput: ~1000 flows/second (single-threaded)
Latency: <10ms per decision
Memory: ~500MB (model + data)

Security Metrics

False Positive Rate: 2.8%
False Negative Rate: 35.9%
ROC AUC (Clean): 0.96
ROC AUC (Adversarial): 0.78 (Degradation analysis)
Adversarial Evasion Success: 20%
Policy Bypass Rate: 0% (context catches evaded flows)

Deployment Architecture

┌─────────────────────────────────────────┐
│         Production Deployment           │
└─────────────────────────────────────────┘

Network Traffic
    ↓
┌─────────────────┐
│ Traffic Mirror  │ ← Span port / TAP
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Feature Extract │ ← Real-time parsing
└────────┬────────┘
         │
         ▼
┌─────────────────────┐
│ Zero-Trust System   │ ← This system
│  (Docker Container) │
└────────┬────────────┘
         │
         ├─→ Access Decision
         │
         ▼
┌─────────────────┐
│ SIEM / SOC      │ ← Splunk, ELK
│  (Telemetry)    │
└─────────────────┘

Future Enhancements

Real-Time Integration
- Live packet capture
- Streaming inference
- Sub-millisecond latency
Advanced ML
- Adversarial training
- Ensemble models
- Online learning
Enhanced Context
- Behavioral analytics
- Historical risk scoring
- Peer group analysis
Explainability
- SHAP values
- Feature importance
- Decision explanations
Automation
- Auto-remediation
- Dynamic policy updates
- Threat intelligence integration

References

NSL-KDD Dataset: UNB CIC
Zero-Trust: NIST SP 800-207
Adversarial ML: ART Framework

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Zero-Trust Network Architecture: Adversarial Attack Detection in ML-Based ZT

System Overview

High-Level Architecture

Component Details

1. Network Data Layer

2. ML Risk Classifier

3. Context Enrichment

4. Zero-Trust Policy Engine

5. Adversarial Attack Layer

FGSM (Fast Gradient Sign Method)

PGD (Projected Gradient Descent)

Data Flow

Normal Flow Processing

Adversarial Flow Processing

Threat Model

Attacker Capabilities

Defense Mechanisms

Layer 1: ML Detection

Layer 2: Zero-Trust Context

Layer 3: Policy Enforcement

System Metrics

Performance Metrics

Security Metrics

Deployment Architecture

Future Enhancements

References

FilesExpand file tree

ARCHITECTURE.md

Latest commit

History

ARCHITECTURE.md

File metadata and controls

Zero-Trust Network Architecture: Adversarial Attack Detection in ML-Based ZT

System Overview

High-Level Architecture

Component Details

1. Network Data Layer

2. ML Risk Classifier

3. Context Enrichment

4. Zero-Trust Policy Engine

5. Adversarial Attack Layer

FGSM (Fast Gradient Sign Method)

PGD (Projected Gradient Descent)

Data Flow

Normal Flow Processing

Adversarial Flow Processing

Threat Model

Attacker Capabilities

Defense Mechanisms

Layer 1: ML Detection

Layer 2: Zero-Trust Context

Layer 3: Policy Enforcement

System Metrics

Performance Metrics

Security Metrics

Deployment Architecture

Future Enhancements

References