Design and Development of an ML/DL Attack Resistance of RC-Based PUF for IoT Security

1. Problem Statement

The Internet of Things (IoT) ecosystem requires lightweight, fast, and reliable device authentication. While Physically Unclonable Functions (PUFs) offer a promising hardware-based solution by leveraging manufacturing variations to generate unique device "fingerprints," they face a critical vulnerability: Machine Learning (ML) and Deep Learning (DL) modeling attacks.

Traditional PUFs (e.g., Ring Oscillator, Arbiter PUFs) often exhibit challenge-response patterns (CRPs) that can be learned by ML algorithms. Attackers can train models on a subset of CRPs to predict responses for unseen challenges with high accuracy, effectively cloning the PUF and compromising security. The paper addresses the need for a PUF architecture that is inherently resistant to these data-driven modeling attacks, particularly for resource-constrained IoT devices where heavy encryption is impractical.

2. Methodology

The authors proposed and evaluated a custom-developed, dynamically reconfigurable Resistor-Capacitor (RC) based PUF. The study followed a structured experimental pipeline:

• Hardware Design:
  • The PUF utilizes simple passive RC components, ensuring ultra-low power consumption and minimal area footprint.
  • Unlike digital timing-race PUFs, this design derives uniqueness from intrinsic analog variability in RC time constants caused by fabrication mismatches.
  • It supports dynamic reconfiguration (First-Order and Second-Order architectures) and operates with 32-bit challenges and 32-bit responses.
• Dataset Generation:
  • A dataset of 80,000 Challenge-Response Pairs (CRPs) was collected under various configurations (different pulse widths, UID integration, and RC orders).
  • The dataset was split into 70% training, 20% validation, and 10% testing.
  • No feature engineering was applied; the data was treated as a "black-box" observation to ensure a realistic attack simulation.
• Attack Simulation (Modeling):
  • The authors tested the PUF's resistance against five distinct ML/DL models:
    1. Decision Trees (DT)
    2. Random Forests (RF)
    3. Extreme Gradient Boosting (XGBoost)
    4. Artificial Neural Networks (ANN)
    5. Gradient Boosted Neural Networks (GBNN)
  • Training Strategy: Models were trained to maximize accuracy (targeting 100% on training data) to ensure they had sufficient capacity to learn the CRPs if a pattern existed.
  • Evaluation Metrics: Bitwise accuracy, Hamming loss, and Exact Match Accuracy (requiring all 32 bits to be correct simultaneously).

3. Key Contributions

• Novel RC-PUF Architecture: Introduction of a dynamically reconfigurable RC-based PUF that leverages analog entropy rather than digital timing races, making the challenge-response mapping non-linear and decorrelated.
• Comprehensive Adversarial Evaluation: A systematic assessment against a wide spectrum of modern ML/DL techniques, ranging from classical ensemble methods (RF, XGBoost) to deep learning (ANN, GBNN).
• Proof of Non-Learnability: Demonstration that even when models achieve perfect training accuracy, they fail to generalize to unseen data, proving the CRPs are not learnable.
• Statistical Validation: Detailed entropy analysis showing that the PUF outputs maintain high entropy (≈0.99), confirming statistical independence and unpredictability.

4. Results

The experimental results highlight a stark contrast between training performance and generalization capability:

• Training Accuracy: All five models achieved ~100% accuracy on the training set, confirming they successfully memorized the training CRPs.
• Test/Validation Accuracy: On unseen test data, all models collapsed to near-random guessing levels:
  • ANN: 51.06%
  • GBNN: 53.27% (highest, but still negligible)
  • XGBoost: 50.97%
  • Decision Tree: 50.37%
  • Random Forest: 51.32%
• Exact Match Accuracy: 0% across all models. No model could predict the full 32-bit response correctly for a single test case.
• Entropy Analysis: The target responses showed near-ideal entropy (0.9926). While most models preserved this, GBNN showed a significant drop (0.67), indicating mode collapse and bias, further proving it failed to learn the true mapping.
• Comparison: Unlike existing PUFs (Arbiter, RO, XOR Arbiter) which often yield 60–90% test accuracy under ML attacks, the proposed RC-PUF maintained test accuracy strictly within the 50–53% range.

5. Significance and Conclusion

The paper concludes that the proposed RC-based PUF offers high resistance to ML/DL modeling attacks. The "failure" of advanced AI models to predict responses beyond random chance is the primary indicator of the system's security.

• Security Implication: The dynamic reconfiguration and analog nature of the RC-PUF create a non-deterministic, high-entropy mapping that prevents attackers from extracting a functional relationship between challenges and responses.
• IoT Relevance: This solution provides a lightweight, energy-efficient, and secure alternative to traditional cryptographic methods for IoT authentication, requiring minimal hardware resources while offering robust protection against next-generation AI-driven threats.
• Future Outlook: The work validates that analog-based, reconfigurable PUF architectures are a viable path forward for securing edge devices against increasingly sophisticated machine learning attacks.