Securing AI/ML Models in IoT Devices: Lessons from Power-Based Adversarial Attacks

As AI/ML models become core components of IoT security systems, their vulnerability to adversarial manipulation presents serious risks. While most adversarial research focuses on image or text data, my work explores a less studied but highly potent vector: power side-channel signals.

This post introduces how attackers can exploit these physical signals to evade AI-based anomaly detection, and outlines strategies for improving robustness.

⚡ Why Power Traces Matter

Modern IoT devices often embed lightweight AI models for malware detection. Instead of analyzing logs or traffic, some advanced approaches monitor power consumption patterns to detect abnormal behavior.

But here’s the problem: these power signatures can be deliberately manipulated.

🧠 My Research: Dummy Code as an Adversarial Tool

In my Ph.D. research, I introduced a runtime dummy code injection technique that creates adversarial perturbations in power signals. These perturbations are:

• Non-intrusive: Do not affect functionality
• Targeted: Designed using SHAP to fool specific model features
• Model-agnostic: Effective against LSTM, CNN, and transformer-based architectures

The result? A 96% misclassification rate across power-based classifiers during Mirai malware’s scanning phase.

🔄 A Multimodal, Transformer-Based Defense Strategy

To counter such attacks, I’m building a multimodal anomaly detection pipeline. It fuses:

• 📊 Static features (PE headers, imports)
• 🔄 Dynamic logs (API calls, network behavior)
• ⚡ Power traces (side-channel)

Using transformer-based models, this architecture learns both temporal and feature relationships while remaining more resilient to perturbations.

🛡️ Security-By-Design: Applying Microsoft Threat Modeling

At HP, I apply Microsoft’s Threat Modeling methodology to assess how attackers could manipulate data, escalate privileges, or spoof communications across edge devices.

For example, what happens if power channels are spoofed to look benign during malicious operations?

These questions help refine AI architectures before deployment, ensuring trust and resilience.

✅ Takeaways

• Power-based detection is powerful but vulnerable
• Multimodal, explainable models offer better protection
• Security must be proactive, with threat modeling integrated into AI design

📚 Learn More

Check out the full project here:
👉 PowerAdversarial-ML

Thanks for reading! I’m always open to collaborations and discussions.
📫 saharzargarzadeh.github.io