TL;DR
This paper introduces the first reconfigurable and ultra-low-power Physical Unclonable Function (PUF) based on Ferroelectric Field-Effect Transistors (FeFETs). By leveraging FeFET's inherent cycle-to-cycle variations and a novel charge-domain in-memory computing approach, the authors achieve remarkable reconfigurability, energy efficiency, and strong resilience against machine learning attacks, making it a promising solution for IoT security.
What Problem Does This Paper Address?
The proliferation of Internet of Things (IoT) devices has created a critical need for robust and lightweight security solutions. Traditional security methods are often too power-hungry or resource-intensive for these edge devices. Physical Unclonable Functions (PUFs) offer a promising alternative, using unique manufacturing variations within a chip to create a hardware-based "fingerprint" for authentication. However, existing PUFs often struggle to balance key requirements: high reconfigurability (to update security keys or adapt to changing threats), ultra-low power consumption, and a vast challenge-response pair (CRP) space to uniquely identify billions of devices. Furthermore, strong PUFs, which offer a large CRP space, are often vulnerable to sophisticated machine learning (ML) attacks. This paper addresses the challenge of creating a PUF that excels in all these areas simultaneously, particularly focusing on reconfigurability and ML attack resilience.
Methods
The core innovation lies in the use of Ferroelectric Field-Effect Transistors (FeFETs) as the fundamental building blocks for the PUF. FeFETs, which integrate a ferroelectric layer into the gate stack of a transistor, offer non-volatility and utilize the polarization state of the ferroelectric material to store information. The authors exploit the cycle-to-cycle (C2C) variation of FeFETs – subtle, random changes in their electrical characteristics over repeated operations – as the primary source of randomness for the PUF. This is a departure from prior work that often relied on device-to-device (D2D) variations, which can be less stable and harder to control for reconfigurability.
A novel 2FeFET-1C PUF cell is proposed, consisting of two FeFETs and a capacitor. This cell is integrated into a larger PUF array and utilizes charge-domain computing-in-memory (CiM). In this approach, computations (like averaging voltages) happen directly within the memory array, eliminating the need for separate processing units and significantly reducing power consumption. The registration and response generation processes are designed to leverage the FeFET C2C variation exclusively, ensuring a homogeneous entropy source for both operations and enabling high reconfigurability. For enhanced ML attack resilience, an XOR PUF structure is also employed, where the outputs of multiple basic PUFs are XORed together to create a more complex and obfuscated response.
Datasets
The paper relies on experimental measurements from fabricated FeFET arrays. No specific external datasets are mentioned for training or evaluation. The performance metrics are evaluated based on the data generated from their fabricated PUF chips.
Key Insights and Ideas
- FeFET C2C Variation as a Superior Entropy Source: The paper convincingly demonstrates that exploiting FeFET's cycle-to-cycle variation, rather than device-to-device variation, leads to more stable and controllable randomness, crucial for high reconfigurability.
- Charge-Domain CiM for Ultra-Low Power: Integrating computation directly into the PUF array using charge-domain principles dramatically slashes readout energy, a critical factor for power-constrained IoT devices.
- All-Around C2C for Reconfigurability: Using C2C variation for both registration and reconfiguration eliminates the negative impacts of D2D variation, leading to near-ideal reconfigurability performance.
- XOR PUF for ML Resilience: Combining the FeFET PUF with an XOR structure significantly boosts resilience against sophisticated machine learning attacks, making it "almost invulnerable."
Key Results
- Near-Ideal Reconfigurability: Achieved an HDreconfig of 50.02%, very close to the ideal 50% for perfect uniqueness between reconfiguration cycles.
- Ultra-Low Readout Energy: Demonstrated an exceptionally low energy consumption of 4.014 fJ/bit, outperforming state-of-the-art designs.
- Excellent Uniformity and Uniqueness: Reported near-ideal uniformity (50.00% HW) and high uniqueness (HDinter of 49.98%).
- 100% Repeatability: Achieved 0% BER (HDintra), indicating perfect repeatability of responses.
- Robustness: Showed high robustness against variations in write voltage, device dimensions, and temperature.
- ML Attack Resilience: The 8-XOR version of the PUF achieved an impressive 52.1% prediction accuracy against various ML models, even with a large training set, making it "almost unpredictable."
Limitations and Future Work
- Readout Energy for XOR PUF: While the basic FeFET PUF has ultra-low energy, the enhanced 8-XOR version for ML attack resilience increases the readout energy to 33.86 fJ/bit. Future work could focus on reducing this energy while maintaining high resilience.
- Scalability and Integration: The paper mentions the potential for macro-level integration due to CMOS compatibility. Further research could explore scaling up the PUF arrays for even larger CRP spaces and integrating them into complex IoT systems.
- ML Attack Configurations: While extensive ML models were tested, future work could delve deeper into specific attack vectors and defenses tailored to this FeFET-based PUF architecture.
Summary updated