What is it about?
This article reviews how researchers test and improve the reliability of TinyML systems. TinyML is a field of artificial intelligence that runs machine-learning models directly on very small, low-power devices, such as microcontrollers, sensors, and wearables. These devices can make decisions close to where data is collected, but they have strict limits in memory, processing power, and energy use. The review examines studies that evaluated whether TinyML models remain reliable when facing real-world challenges such as sensor noise, changing environments, adversarial attacks, hardware faults, limited resources, data imbalance, and new deployment conditions. The article also maps which application areas have been studied, what datasets and input types are commonly used, and which metrics researchers rely on to measure robustness.
Featured Image
Photo by Arturo Esparza on Unsplash
Why is it important?
TinyML is increasingly used in practical settings such as healthcare monitoring, robotics, cybersecurity, environmental sensing, and human-machine interaction. In these areas, a model that works well in the lab may fail when deployed on a real device exposed to noise, limited battery life, hardware imperfections, or changing data. This review is important because it shows that reliability in TinyML is not only about achieving high accuracy. It also depends on whether models can remain stable under difficult conditions while respecting memory, energy, and latency constraints. By organizing existing research into clear robustness categories and application areas, the article helps researchers and practitioners understand what has already been studied, what remains underexplored, and how future TinyML systems can be made more dependable, secure, and deployment-ready.
Perspectives
As an author, I see this work as a step toward making TinyML research more useful for real-world deployment. TinyML has great potential because it can bring artificial intelligence to small, low-cost, and energy-efficient devices, enabling faster decisions, reduced dependence on cloud services, and better privacy. However, these benefits only matter if the models remain reliable once they leave controlled experimental settings. Through this review, we wanted to show where the field stands today and where stronger evaluation practices are still needed. I hope this article encourages the community to evaluate TinyML models not only by accuracy, but also by robustness, memory use, energy cost, hardware behavior, and practical deployment constraints. In the long term, this can help build embedded AI systems that are safer, more transparent, and more trustworthy.
Emanuel Pereira
Universidade Federal do Rio Grande do Norte
Read the Original
This page is a summary of: Robustness in TinyML: A Systematic Literature Review, ACM Transactions on Embedded Computing Systems, June 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3820656.
You can read the full text:
Contributors
The following have contributed to this page
