Deep PUF: A Highly Reliable DRAM PUF-Based Authentication for IoT Networks Using Deep Convolutional Neural Networks.

Area of Science:

• Computer Science
• Cybersecurity
• Hardware Security

Background:

• Traditional cryptographic authentication methods are vulnerable to attacks on session keys and data.
• Physical Unclonable Functions (PUFs), specifically DRAM-based PUFs, offer a promising alternative for secure authentication.
• Existing PUF implementations often suffer from reliability issues due to noise, necessitating complex error-reduction techniques and increasing overhead.

Purpose of the Study:

• To propose a novel authentication scheme, Deep PUF, leveraging deep convolutional neural networks (CNNs) and latency-based DRAM PUFs.
• To eliminate the need for additional error correction codes (ECCs) or pre-selection schemes in PUF-based authentication.
• To enhance the number of challenge-response pairs (CRPs) and enable authentication for resource-constrained devices by offloading complexity to the server.

Main Methods:

• Implementation of a deep convolutional neural network (CNN) for classifying responses from latency-based DRAM PUFs.
• Utilizing a 1Gb DDR3 memory module for experimental validation.
• Developing an authentication framework that integrates CNN classification results to minimize identification errors.

Main Results:

• The proposed CNN model achieved a classification accuracy of at least 94.9% for DRAM PUF responses under varying conditions.
• The Deep PUF scheme successfully eliminated the need for traditional error-reduction methods like ECCs and pre-selection.
• Experimental results demonstrated a drastic reduction in the probability of identification error, leading to highly reliable authentication.

Conclusions:

• Deep PUF offers a robust and reliable authentication solution for devices, particularly those with limited resources.
• The integration of CNNs with DRAM PUFs significantly improves authentication security and efficiency.
• This approach overcomes the inherent reliability challenges of PUFs without introducing substantial overhead.