Securing wireless communications in harsh environments, such as underwater networks, via traditional cryptographic approaches is unfeasible. For example, public key encryption would require a public key infrastructure and a key management infrastructure. A viable solution is instead physical layer security, allowing two devices to obtain a symmetric cryptographic key from the randomness provided by the underlying communication channel, which varies in time, frequency, and space, in general. The probability of having both parties generating the same key and its number of bits greatly depend on how sampled observations are quantized. In this thesis, novel datadriven quantization techniques, which make use of specific channel features computed from impulse responses collected from real experiments, are investigated. In particular, we propose a new machine learning algorithm that quantizes an input vector into an initial key, as close as possible to a series of independent and uniformly distributed symbols and matches at beast the corresponding initial key of the corresponding receiver, to guarantee a high key agreement probability and to avoid an eavesdropper to infer future values exploiting the correlation between consecutive symbols. We also propose an adversarial neural network architecture, where legitimate parties feature a neural quantizer to produce the initial key, whereas the eavesdropper tries to reconstruct the key agreed by the first two.
Quantization for Secret Key Generation in Underwater Acoustic Channels
GIULIANI, AMEDEO
2021/2022
Abstract
Securing wireless communications in harsh environments, such as underwater networks, via traditional cryptographic approaches is unfeasible. For example, public key encryption would require a public key infrastructure and a key management infrastructure. A viable solution is instead physical layer security, allowing two devices to obtain a symmetric cryptographic key from the randomness provided by the underlying communication channel, which varies in time, frequency, and space, in general. The probability of having both parties generating the same key and its number of bits greatly depend on how sampled observations are quantized. In this thesis, novel datadriven quantization techniques, which make use of specific channel features computed from impulse responses collected from real experiments, are investigated. In particular, we propose a new machine learning algorithm that quantizes an input vector into an initial key, as close as possible to a series of independent and uniformly distributed symbols and matches at beast the corresponding initial key of the corresponding receiver, to guarantee a high key agreement probability and to avoid an eavesdropper to infer future values exploiting the correlation between consecutive symbols. We also propose an adversarial neural network architecture, where legitimate parties feature a neural quantizer to produce the initial key, whereas the eavesdropper tries to reconstruct the key agreed by the first two.File  Dimensione  Formato  

giuliani_amedeo.pdf
accesso aperto
Dimensione
9.35 MB
Formato
Adobe PDF

9.35 MB  Adobe PDF  Visualizza/Apri 
The text of this website © Università degli studi di Padova. Full Text are published under a nonexclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/35582