PhyDNNs: Bringing Deep Neural Networks to the Physical Layer

Emerging applications require mobile devices to continuously execute complex deep neural networks (DNNs). While mobile edge computing (MEC) may reduce the computation burden of mobile devices, it exhibits excessive latency as it relies on encapsulating and decapsulating frames through the network protocol stack. To address this issue, we propose PhyDNNs, an approach where DNNs are modified to operate directly at the physical layer (PHY), thus significantly decreasing latency, energy consumption, and network overhead. Conversely from recent work in Joint Source and Channel Coding (JSCC), PhyDNNs adapt already trained DNNs to work at the PHY. To this end, we developed a novel information-theoretical framework to fine-tune PhyDNNs based on the trade-off between commu-nication efficiency and task performance. We have prototyped PhyDNNs with an experimental testbed using a Jetson Orin Nano as the mobile device and two USRP software-defined radios (SDRs) for wireless communication. We evaluated PhyDNNs performance considering various channel conditions, DNN models, and datasets. We also tested PhyDNNs on the Colosseum net-work emulator considering two different propagation scenarios. Experimental results show that PhyDNNs can reduce the end-to-end inference latency, amount of transmitted data, and power consumption by up to 48×, 1385×, and 13 × while keeping the accuracy within 7 % of the state-of-the-art approaches. Moreover, we show that PhyDNNs experience 4.3 times less latency than the most recent JSCC method while incurring in only 1.79% performance loss. For replicability, we shared the source code for the PhyDNNs implementation.