Proceedings of the IE 2020 International Conference www.conferenceie.ase.ro 14 A.I. NEURAL NETWORKS INFERENCE INTO THE IOT EMBEDDED DEVICES USING TINYML FOR PATTERN DETECTION WITHIN A SECURITY SYSTEM Cristian TOMA Department of Economic Informatics & Cybernetics Bucharest University of Economic Studies, Romania cristian.toma@ie.ase.ro Marius POPA Department of Economic Informatics & Cybernetics Bucharest University of Economic Studies, Romania marius.popa@ie.ase.ro Mihai DOINEA Department of Economic Informatics & Cybernetics Bucharest University of Economic Studies, Romania mihai.doinea@ie.ase.ro Abstract. The paper presents the implementation challenges of a proof of concept development for image processing with artificial intelligence neural network into an embedded device. Because of the hardware constraints of the embedded device – e.g. 16 KB RAM and 128 KB of EEPROM, the deep learning/the training and the model of the neural network is processed in cloud – e.g. Google Cloud Platform – AI – Artificial intelligence, as shown in second section. After this step, the trained neural networks model, values and tensors are translated with TinyML into native code for the embedded devices and deployed on a specific hardware platform – e.g. Arduino Nano 33 BLE Sense or SparkFun Edge Development Board Apollo3 Blue - for the neural network inferences – e.g. person detection or NLP into specific area, as described in third section. The last section shows the conclusions and the security challenges for deploying neural networks into embedded systems which are used for security systems such as: monitoring and surveillance cameras, drones visual computing for securing field areas, IoT systems, etc. Keywords: Artificial Intelligence, Neural Networks, Deep Learning, embedded devices, TinyML, cybersecurity, IoT. JEL classification: C88, L86, Y80 DOI: 10.24818/ie2020.01.03 1. Introduction into Neural Networks Inferences Process There are multiple software development kits, libraries and frameworks for developing Neural Networks and Deep Learning applications. Most of them are pushing Python based code, but Python has poor results for the performance benchmarking on real CPU and GPU cores. This maybe an advantage for the cloud providers but not necessary an advantage for the companies or software integrators who are renting processing power into the Cloud. Most used framework/API is Keras. are which can be accessible from the Keras API. Keras API may address via API multi-back-end Artificial Intelligence Cloud solutions, such as: Tensorflow, Theano, MxNet. Tensorflow is polyglot and is supporting Java and JVM based languages as well, but for the moment, Python is pushed into a lot of books and tutorials. Additionally, in