Improving IoT security through the use of deep learning at the physical layer

The Internet of Things (IoT) is a heterogeneous network interconnection connecting electronic and electro-mechanical devices to the Internet. The total number of IoT devices is estimated to reach 26.66 billion and is expected to reach 75.4 billion by 2025. Currently, only 30% of the IoT devices employ encryption, which puts the majority of the IoT devices and their underlying infrastructure under risk of attacks by: 1) devices that are wrongly authenticated to access the network specially when digital credentials are transmitted without encryption, and 2) devices that can detect, intercept, and exploit communications between IoT devices. Therefore, more advanced security mechanism are required to secure IoT devices, their corresponding networks, and infrastructure. The Open Systems Interconnect (OSI) stack provides a layered model that governs IoT networks. Based on the OSI stack, the physical (PHY) layer–of each IoT device and associated network–is the first layer exposed to attacks. Traditionally, IoT security techniques are implemented in higher OSI layers, thus these techniques ignore the PHY layer and any potential security advantages it possesses. Due to the demonstrated success of Deep Learning (DL) within the fields of computer vision and image processing, as well as prior research that suggests DL as a viable solution to addressing communications system challenges; this work investigates DL-driven PHY layer security techniques that surpass traditional approaches. The presented work investigates PHY layer security at the encoding and waveform levels. Encoding-based PHY layer security is achieved through an adversarial training and shared-code scheme that leverages DL to redesign a Direct Sequence Spread Spectrum (DSSS) communications system such that it inherently, deliberately, and adaptively prevents an adversary from detecting and reconstructing captured messages. Waveform based PHY layer security is improved through a Radio Frequency-Distinct Native Attributes (RF-DNA) fingerprint process capable of exploiting Specific Emitter Identification (SEI) features that are extracted from waveforms that transverse a Rayleigh fading channel prior to collection. This is achieved through the integration of channel correction prior to DL-based radio identification. The investigated channel correction approaches include traditional and semi supervised learning. The results show that: 1) the DL-based redesign of DSSS encoding achieves featureless signaling that prevents the adversary from reconstructing detected messages, and 2) unsupervised learning based channel correction improves RF-DNA fingerprinting performance by 25% over that of traditional machine learning approaches