IoT Security Case Study: Edge AI & Federated Learning in Smart Grids

 1Timeline:
 2  Dec 2015: BlackEnergy malware infects 3 power distribution companies
 3  Attack Flow:
 4    1. Spear-phishing → Weak credential compromise
 5    2. Lateral movement → SCADA network via jump host
 6    3. Remote code execution → Firmware installation
 7    4. Synchronized disconnect → 230,000+ customers without power
 8  Duration: ~2 hours
 9  Impact: $43.2M in emergency restoration costs
10  Root Cause: Unpatched Windows vulnerability + phishing

1Hypothetical PMU Compromise:
2  1. Firmware update trojanized → PMU receives false frequency readings
3  2. PMU reports 50.5 Hz (nominal: 50.0 Hz)
4  3. Grid operators see apparent overfrequency → Shed load
5  4. Cascade failure → Regional blackout
6  5. Actual frequency: 49.8 Hz (underfrequency) → Generator damage
7  Consequence: Days to restore; $billions in economic loss

1Example Attack:
2  Attacker queries: "Customers with >500 kWh/week overnight?"
3  Response: List includes data centers, cryptocurrency miners, factories
4  Attacker: Plan physical theft during high-activity period (easier evasion)
5  
6  Or: "When are industrial facilities on 4-day weekend?"
7  Response: Identify burglary targets based on occupancy patterns

 1┌──────────┐   ┌──────────┐   ┌──────────┐
 2│ Meter 1  │─→ │ Meter 2  │─→ │ Meter N  │
 3│ 500kWh   │   │ 300kWh   │   │ 250kWh   │
 4└──────────┘   └──────────┘   └──────────┘
 5        │               │               │
 6        └───────────────┴───────────────┘
 7                     │
 8          ┌──────────▼──────────┐
 9          │ Central Server      │
10          │ All raw data        │
11          │ Vulnerable target!  │
12          └─────────────────────┘

 1Local Edge                     Aggregation                Server
 2─────────────────────────────────────────────────────────────
 3
 4Meter 1                        ┌─────────────┐
 5├─Train model locally          │ Secure      │
 6├─Add DP noise                 │ Aggregation │
 7├─Send update (not raw data)   │ (Threshold  │  ┌───────────┐
 8│                              │  Crypto)    │─→│ Global    │
 9Meter 2                        │             │  │ Model     │
10├─Train model locally          │ Combine     │  └───────────┘
11├─Add DP noise                 │ weights     │
12├─Send update                  │             │
13│                              │ Only model  │
14Meter N                        │ updates     │
15├─Train model locally          │ transmitted │
16├─Add DP noise                 └─────────────┘
17├─Send update
18│
19No raw data leaves meter!

 1ε-δ Differential Privacy Definition:
 2  For datasets D and D' differing by 1 customer:
 3  P(M(D) ∈ S) ≤ e^ε · P(M(D') ∈ S) + δ
 4  
 5Intuition:
 6  Small ε → Strong privacy (query result same whether customer included or not)
 7  ε = 0.1 → Attacker gains <11% confidence over random guess
 8  ε = 1.0 → Attacker gains ~2.7x advantage (privacy still strong)
 9  
10Privacy Budget:
11  Each query costs ε
12  Total budget = 1.0 → After ~10 queries, exhausted
13  Force ε decay over time (stronger privacy for later queries)

1def local_dp_gradient(gradient, sensitivity, epsilon):
2    """Add Laplace noise for local DP"""
3    scale = sensitivity / epsilon
4    noise = np.random.laplace(0, scale, size=gradient.shape)
5    return gradient + noise

 1Protocol (Bonawitz et al., 2017):
 2  Setup Phase:
 3    Server generates public key pairs for each client
 4    Each client encrypts local model updates
 5  
 6  Aggregation Phase:
 7    Server sums encrypted values: E(Σ updates) 
 8    Server never sees unencrypted individual updates
 9    Only decrypts final sum
10  
11  Dropout Handling:
12    If client disappears, protocol restarts
13    All remaining clients re-encrypt (security preserved)

1Traditional: Meter → Server (raw consumption) → Decision
2            Exposure window: Data in transit
3
4Federated+Edge: Meter → Local inference → Decision
5               Exposure window: Eliminated

1Asset Class     |  Value    | Attack Impact           | Criticality
2─────────────────────────────────────────────────────────────────
3Customer Data   |  PII      | Privacy violation       | HIGH
4Grid Topology   |  Strategic| Targeted attacks        | CRITICAL
5Control Logic   |  Strategic| Malware injection       | CRITICAL
6Meter Integrity |  $500-2K  | Revenue loss (kWh fraud)| MEDIUM

 1                Impact
 2        L      M      H      C
 3      ┌────┬────┬────┬────┐
 4    L │ Low│Low │Med│High│
 5      ├────┼────┼────┼────┤
 6C  M │Low │Med │Med│High│
 7r    ├────┼────┼────┼────┤
 8i  H │Med │Med │High│Crit│
 9t    ├────┼────┼────┼────┤
10    C │High│Crit|Crit|Crit│
11      └────┴────┴────┴────┘
12Likelihood →

 1                 ┌──────────────────┐
 2                 │ User Awareness   │
 3                 │ Security Training│
 4                 └────────┬─────────┘
 5                          │
 6                 ┌────────▼─────────┐
 7                 │ Endpoint Security │
 8                 │ EDR, Antivirus   │
 9                 └────────┬─────────┘
10                          │
11                 ┌────────▼─────────┐
12                 │ Network Security │
13                 │ Segmentation     │
14                 │ DDoS Protection  │
15                 └────────┬─────────┘
16                          │
17                 ┌────────▼─────────┐
18                 │ OT/SCADA Security│
19                 │ Air-gap critical │
20                 │ systems          │
21                 └────────┬─────────┘
22                          │
23                 ┌────────▼─────────┐
24                 │ Data Privacy     │
25                 │ Encryption at-rest
26                 │ DP + FL          │
27                 └──────────────────┘

 1                    Internet
 2                       │
 3                       │ (Encrypted, Authenticated)
 4                       │
 5            ┌──────────┴──────────┐
 6            │ Edge Gateway        │
 7            │ (Aggregation point) │
 8            │ DP application      │
 9            └──────────┬──────────┘
10                       │ Local network
11           ┌───────────┼───────────┐
12           │           │           │
13      ┌────▼──┐    ┌────▼──┐  ┌────▼──┐
14      │Meter 1│    │Meter 2│  │Meter N│
15      │(DP)   │    │(DP)   │  │(DP)   │
16      └───────┘    └───────┘  └───────┘

 1# Simulated FL system for smart grid
 2
 3class FederatedGridOptimizer:
 4    def __init__(self, epsilon=1.0, num_clients=100):
 5        self.epsilon = epsilon
 6        self.global_model = initialize_model()
 7        
 8    def local_update(self, client_id, local_data):
 9        """Client-side: Train locally + add noise"""
10        model = self.global_model.copy()
11        
12        # Train on local data (meter consumption)
13        for epoch in range(5):
14            loss = model.train_step(local_data)
15        
16        # Extract gradients
17        gradients = model.get_gradients()
18        
19        # Add Laplace noise for DP
20        sensitivity = estimate_gradient_sensitivity()
21        noisy_gradients = self._apply_laplace_noise(
22            gradients, sensitivity, self.epsilon
23        )
24        
25        return noisy_gradients
26    
27    def aggregate(self, client_updates):
28        """Server-side: Aggregate without seeing raw updates"""
29        avg_update = np.mean(client_updates, axis=0)
30        self.global_model.apply_update(avg_update)
31        
32    def _apply_laplace_noise(self, gradient, sensitivity, epsilon):
33        scale = sensitivity / epsilon
34        noise = np.random.laplace(0, scale, size=gradient.shape)
35        return gradient + noise

 1class ComplianceMonitor:
 2    def __init__(self, nerc_cip_rules):
 3        self.rules = nerc_cip_rules
 4        self.violations = []
 5        
 6    def audit_configuration(self, system_config):
 7        """Check against NERC CIP requirements"""
 8        for rule in self.rules:
 9            if not rule.verify(system_config):
10                self.violations.append({
11                    'rule': rule.id,
12                    'severity': rule.severity,
13                    'remediation': rule.remediation
14                })
15        return self.violations

 1Query: "Average consumption in neighborhood X?"
 2
 3Without DP:
 4  Attacker: (10 queries) → Reconstruct individual consumption
 5  Success: ~85% accuracy
 6
 7With DP (ε=1.0):
 8  Attacker: (10 queries) → Noisy estimates
 9  Success: ~5% accuracy (worse than random)
10  Privacy: **PROTECTED**

1Metric              | Centralized | Federated+DP | Overhead
2─────────────────────────────────────────────────────────────
3Model Accuracy      | 96.2%       | 94.8%        | -1.4pp
4Training Time       | 4.2 hours   | 2.1 hours    | -50% (parallel)
5Privacy Guarantee   | None        | ε=1.0 (strong) | +0%
6Data Transmission   | 500GB/day   | 50MB/day     | -99.99%
7Scalability (1K→10K nodes) | 2.1× slowdown | 1.05× slowdown | **Pass**

 1# Setup federated learning environment
 2git clone https://github.com/[user]/smart-grid-fl
 3cd smart-grid-fl
 4
 5# Install dependencies
 6pip install tensorflow-federated opacus
 7
 8# Run simulation
 9python federated_grid_optimizer.py \
10  --num_clients=100 \
11  --epsilon=1.0 \
12  --epochs=10
13
14# Generate compliance report
15python compliance_audit.py --standard=nerc_cip

 1┌─────────────────────────────────────────┐
 2│         Smart Meter Devices              │
 3│  ┌──────────────────────────────────┐   │
 4│  │ Local Federated Learning Model   │   │
 5│  │ + Differential Privacy Noise     │   │
 6│  └──────────────────────────────────┘   │
 7└──────────────┬──────────────────────────┘
 8               │ Encrypted Model Updates
 9               ▼
10┌─────────────────────────────────────────┐
11│    Secure Aggregation Service            │
12│  (Threshold Cryptography)                │
13│  No plaintext data access                │
14└──────────────┬──────────────────────────┘
15               │ Aggregate Model
16               ▼
17┌─────────────────────────────────────────┐
18│       Utility Control Center              │
19│  (Can use model for anomaly detection)   │
20│  (Cannot identify individual users)      │
21└─────────────────────────────────────────┘

1# View paper
2open case_study_abs.pdf
3
4# Access supporting materials
5cd group_2_question/   # Research questions
6cd group_2_ppt/        # Presentation slides

 1Case_study/
 2├── case_study_abs/
 3│   ├── case_study_abs.tex
 4│   ├── case_study_abs.pdf
 5│   ├── references.bib
 6│   └── case_study_abs.aux
 7├── case_study_one_page/ (Extended abstract)
 8├── group_2_ppt/ (Presentation slides)
 9├── group_2_question/ (Research questions)
10└── readme.md