Abstract Physical Unclonable Functions (PUFs) are novel circuit structures that provide hardware security solutions in application areas such as chip design and IoT, due to characteristics of their lightweight, key-free and tamper-resistant. PUFs are not immune to threats like machine learning modeling attacks and side channel modeling attacks. Strong PUFs are susceptible to classical machine learning attacks, however, machine learning’s effectiveness in attacking complex structured strong PUFs is limited, and its efficiency is relatively low. Side-channel modeling attacks, on the other hand, incur high implementation costs. Hence, employing deep learning for modeling attacks becomes an effective and cost-efficient choice when attacking complex structured PUFs. In this paper, we introduce a method that employs deep neural network to assess the modeling resilience of combination logic operation-based PUFs with APUFs as components for the first time. We employed a 4-layer DNN model to investigate the security resilience of PUF models involving any combination of OR AND and XOR logical operations. We explored the security regular patterns of modeling resilience. We have demonstrated for the first time that bias in PUF responses can reduce or destroy the security of PUFs. OR or AND logic operations do not provide any security benefit in PUF design, while XOR operations enhance the security of PUFs.

Physical unclonable functions (PUFs) have emerged as potent hardware primitives owing to their intrinsic properties of being secret key-free, clone-proof, and lightweight. However, PUFs cannot avoid the threats of machine learning modeling and side-channel attacks (SCAs). Nevertheless, almost all attacks neglect the correlations between the mathematical model and side-channel models introduced by PUF internal parameters; thus, such attacks fail to exploit related data and struggle in modeling complex PUFs. To address this problem, we propose a framework for using auxiliary learning SCAs to model strong PUFs by learning multiple related tasks together. Side-channel information predictions are introduced as auxiliary tasks to facilitate the primary task of predicting response. The parameters hard for the primary task to learn can be shared by the auxiliary tasks that learn the same parameters more straightforwardly. Based on the proposed framework, we design a specific auxiliary learning power SCA that employs power level prediction as the auxiliary task. The proposed attack is implemented with the hard-parameter sharing and hierarchy sharing deep neural networks. Experimental results demonstrate that the proposed attack succeeds in modeling XOR APUF, MPUF, and iPUF and outperforms the state-of-the-art methods in modeling MPUF and iPUF. We evaluate the influences of task relatedness, architecture, and loss weight ratio. Furthermore, we propose a fine-grained classification-based method to generate the auxiliary task with an enhanced relationship to the primary task. According to the response, the class corresponding to a specific side-channel state is further divided into two subclasses. Experimental results demonstrate that the generated auxiliary task promotes performance and alleviates the adverse effects of improper architecture and parameters.