Evaluating the sensitivity of circuits to Single Event Upset (SEU) faults has become increasingly important and challenging due to the growing complexity of circuits. Simulation-based fault injection is time-intensive, particularly for highly complex circuits. This paper proposes a novel approach using Spatio-temporal Graph Convolutional Networks (STGCN) to predict SEU fault propagation results in circuits. By representing circuits’ structure as graphs and integrating temporal features from the simulation workload, STGCNs can learn from these spatio-temporal graphs to identify SEU fault propagation patterns. To validate this method, we test it on six evaluation circuits, achieving a prediction accuracy of 93-99%. Given this performance, to accelerate SEU simulation-based fault injection, we divide SEU faults into three subsets and use a STGCN fine-tuned on the training and validation dataset to predict SEU fault propagation in the test dataset, eliminating the need for simulation and reducing the required time. To identify an efficient dataset separation method, we compare three sampling methods: spatial sampling (sampling flip-flops for injected faults), temporal sampling (sampling time points for fault injection), and hybrid sampling (incorporating both spatial and temporal sampling). The hybrid sampling approach is the most promising, optimizing the trade-off between efficiency and accuracy. This approach reduces simulation time by 50% while maintaining accuracy above 95% on the six evaluation circuits.

Recently, several research works have emphasized the problem of stealing the intellectual property of trained Machine Learning (ML) models from hardware neural network inference engines spotlighting Binarized Neural Networks (BNNs). The binary operations in BNNs can be executed bitwise, which notably saves storage memory, reduces the execution time and power and, therefore, makes them convenient for implementation in hardware. Unfortunately, these advantages may also enable a vulnerability to Differential Power Analysis (DPA) side-channel attacks, which, in turn, necessitates dedicated masking techniques to protect the models. Notably, the recent BNN hardware inference engines are being increasingly adopted for critical applications and demand, along with security, also high levels of in-filed reliability throughout their lifetime. The state-of-the-art power side-channel masking in BNNs implies glitch-resistant structures, such as Trichina AND gates and sequences of flip-flops, and may create soft-error reliability issues that are currently overlooked in the literature. This paper presents an analysis for the soft-error reliability jeopardy by the security countermeasures in hardware implementations of BNN inference engines. Our work reveals a steep increase (hundreds of times) of vulnerability to single-event effects, introduced by the state-of-the-art security enhancement techniques, and emphasizes the interdependency of the design’s reliability and security aspects.

Typical design flows are hierarchical and rely on assembling many individual technology elements from standard cells to complete boards. Providers use compact models to provide simplified views of their products to their users. Designers group simpler elements in more complex structures and have to manage the corresponding propagation of reliability and functional safety information through the hierarchy of the system, accompanied by the obvious problems of IP confidentiality, possibility of reverse engineering and so on. This paper proposes a machine-learningbased approach to integrate the many individual models of a subsystem's elements in a single compact model that can be reused and assembled further up in the hierarchy. The compact models provide consistency, accuracy and confidentiality, allowing technology, IP, component, sub-system or system providers to accompany their offering with high-quality reliability and functional safety compact models that can be safely and accurately consumed by their users.

Selective mitigation or selective hardening is an effective technique to obtain a good trade-off between the improvements in the overall reliability of a circuit and the hardware overhead induced by the hardening techniques. Selective mitigation relies on preferentially protecting circuit instances according to their susceptibility and criticality. However, ranking circuit parts in terms of vulnerability usually requires computationally intensive fault-injection simulation campaigns. This paper presents a new methodology which uses machine learning clustering techniques to group flip-flops with similar expected contributions to the overall functional failure rate, based on the analysis of a compact set of features combining attributes from static elements and dynamic elements. Fault simulation campaigns can then be executed on a per-group basis, significantly reducing the time and cost of the evaluation. The effectiveness of grouping similar sensitive flip-flops by machine learning clustering algorithms is evaluated on a practical example.Different clustering algorithms are applied and the results are compared to an ideal selective mitigation obtained by exhaustive fault-injection simulation.

The Functional Failure Rate analysis of today's complex circuits is a difficult task and requires a significant investment in terms of human efforts, processing resources and tool licenses. Thereby, de-rating or vulnerability factors are a major instrument of failure analysis efforts. Usually computationally intensive fault-injection simulation campaigns are required to obtain a fine-grained reliability metrics for the functional level. Therefore, the use of machine learning algorithms to assist this procedure and thus, optimising and enhancing fault injection efforts, is investigated in this paper. Specifically, machine learning models are used to predict accurate per-instance Functional De-Rating data for the full list of circuit instances, an objective that is difficult to reach using classical methods. The described methodology uses a set of per-instance features, extracted through an analysis approach, combining static elements (cell properties, circuit structure, synthesis attributes) and dynamic elements (signal activity). Reference data is obtained through first-principles fault simulation approaches. One part of this reference dataset is used to train the machine learning model and the remaining is used to validate and benchmark the accuracy of the trained tool. The presented methodology is applied on a practical example and various machine learning models are evaluated and compared.