MADel0: A Modelling and Assessment Framework for Delay PUFs leveraging Gradient-based Optimization Techniques

Physically Unclonable Functions leverage manufacturing variations to generate unique, unclonable responses, making them a critical hardware root of trust. Strong PUFs, with exponentially large challenge-response pairs, are vulnerable to machine learning (ML)-based modeling attacks. Among these, reliability-based attacks pose a serious threat by exploiting response instability under varying conditions to improve model accuracy. Existing approaches rely on standard ML techniques, often overlooking PUF-specific design characteristics. This work redefines reliability-based PUF modeling as an optimization problem and introduces MADelO, a framework employing adaptive gradient descent (GD) with tailored objective functions that integrate both CRPs and reliability information. We formally analyze the optimization objectives and provide convergence guarantees. We study how different PUF design strategies influence convergence rates, offering insights into enhancing resilience against such attacks. To the best of our knowledge, this is the first formal framework for reliability-based modeling attacks. Our results demonstrate the effectiveness of optimization-based approaches, opening new directions of attack strategies.