Researchers at Qatar University, University of Idaho and Temple University, led by Dr. Heena Rathore, are mimicking how the human brain works to prevent low-power IoT devices, such as insulin pumps, brain simulators, cardiac defibrillators and others, used in medical applications from being hacked. They have built a system that replicates how higher organisms evolve the capability for perceptual recognition, generalization, recall and thinking, by learning from the answers to the questions asked by F.  Rosenblatt, in his seminary work titled “Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain”, published in Psychological Review, in 1958. These three questions are related to how biological systems sense information about the physical world, how they store this information, and finally how they use stored information to influence recognition and behavior.

The seminary work done by Rosenblatt is the foundation of the technique, known as Multi-layer Perceptron, widely used today in the specific area of Deep Learning techniques. Deep Learning is a field that researchers use to show how basic models of human brains can be utilized to solve difficult computational tasks like predictive modeling. Neural networks are capable of learning and representing any sensory information and then relating it to the output variables that need to be predicted. This requires storage of such information in the form of coded representations or images and a one-to-one mapping between the sensory stimulus and the stored pattern. According to this hypothesis, if a system is designed to understand the wiring diagram of the underlying neurons, one should be able to discover exactly what an organism remembers by reconstructing the original sensory patterns from the "memory traces" which they have left.

The predictive capacity of neural systems originates from the various single or multi-layered structures of the systems. As shown in Figure 1, a typical building block of neural network systems is artificial neurons.

Figure 1.  A multi-layer deep learning neural network and comparative results.

These are basic computational units that have weighted input information and deliver an output utilizing an activation function. In a multi-layer neural network, the overall computation comprises of two phases, spread across multiple layers. These phases are called forward and backward propagation phases. During the forward propagation, output is predicted on the basis of repetitive weighted input computation and activation function followed by backward propagation. Gradient descent method, a type of optimization process, is used to adjust all the weights in the network by propagating the total error between the output and input nodes. A deep learning based neural network may have multiple layers of hidden nodes, with each layer engaging in a forward and backward propagation phase, with adjusted weights being applied to each phase.

Dr. Rathore has applied this concept to achieving security in low power IoT devices, specifically targeted at wireless medical devices and applications. By implementing these neural networks on a software platform, consisting of Python, Keras and Theano backend, she has shown that this technique achieves more than 90% accuracy and is order of magnitude more efficient than other state of the art techniques, such as Support Vector Machine (SVM), biometric based approaches, distance/proximity-based techniques, key management protocols, audit mechanisms, anomaly detection mechanisms and external device methodologies. This work was presented at Globecom 2017, in a paper titled "DLRT: Deep Learning for Reliable Diabetic Treatment". By performing many of the complex math operations on the device itself, rather than pushing data to the cloud, Dr. Rathore has been able to make these devices more accessible and less susceptible to hacking through the wireless communication medium.

References
1. Rosenblatt, F. (1958). The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review, 65(6), 386-408.
2. H. Rathore, A. Al-Ali, A. Mohamed, X. Du and M. Guizani, "DLRT: Deep Learning Approach for Reliable Diabetic Treatment," GLOBECOM 2017 - 2017 IEEE Global Communications Conference, Singapore, 2017, pp. 1-6.