Robustness in TinyML: A Systematic Literature Review

TinyML enables the deployment of machine learning models on low-power embedded devices, offering energy-efficient solutions for real-world applications. However, TinyML faces significant challenges due to strict memory, processing, and energy constraints, making the implementation of robust and scalable models particularly difficult. Robustness in this context refers to the ability of models to maintain stable performance under adversarial conditions, sensor noise, and environmental variability, which makes it an essential requirement for reliable deployment in practical scenarios. This study conducts a systematic literature review to examine how robustness is assessed in TinyML, analyzing key factors such as input data types, accessibility of the dataset, real vs. simulated data usage, application domains, evaluated robustness types, hardware constraints, and commonly used performance metrics. The findings show a strong preference for sensor-based inputs, public and real-world datasets, and a focus on noise-related robustness challenges. Memory efficiency stands out as the main hardware constraint, while accuracy is the most used evaluation metric, reflecting the dominance of classification tasks in TinyML research. These insights provide a structured overview of current trends and reveal key gaps in TinyML robustness research, such as the lack of standardized benchmarking frameworks and the need for more advanced adversarial defense mechanisms tailored to low-power environments. They also highlight opportunities for integrating federated and physics-informed learning approaches, promoting the development of more secure, efficient, and resilient embedded machine learning systems.