Diffusion Model-Based Assisted Attacks on Pufsecured Telematics and Medical Devices

In this paper, we present a novel approach leveraging Denoising Diffusion Probabilistic Models (DDPMs) to enhance attacks on physically unclonable function (PUF) keys. PUFs generate secret binary keys that are electromagnetically unclonable and have demonstrated robustness against machine learning-based attacks, including Fourier regression and generative adversarial networks (GANs). Our method utilizes DDPMs to generate keys from pure noise or noisy signals, effectively replicating the secret keys extracted from integrated circuits. To optimize performance, we introduce modifications to the pre-processing and postprocessing stages. Specifically, the pre-processing step transitions the key representation from binary to grayscale before applying DDPM, while the post-processing stage incorporates a predefined threshold to binarize the generated keys. Experimental results demonstrate that our DDPM-based approach significantly outperforms GANs, achieving superior accuracy in key generation from noise. This work highlights the potential of DDPMs in assisted attacks by investigating noisy signals.