Optical Processing Units (OPUs) -- low-power photonic chips dedicated to large scale random projections -- have been used in previous work to train deep neural networks using Direct Feedback Alignment (DFA), an effective alternative to backpropagation. Here, we demonstrate how to leverage the intrinsic noise of optical random projections to build a differentially private DFA mechanism, making OPUs a solution of choice to provide a \emph{private-by-design} training. We provide a theoretical analysis of our adaptive privacy mechanism, carefully measuring how the noise of optical random projections propagates in the process and gives rise to provable Differential Privacy. Finally, we conduct experiments demonstrating the ability of our learning procedure to achieve solid end-task performance.

This paper analyzes the impact on the stability of the TCP-Reno congestion control mechanism when a network coding (NC) layer is inserted in the TCP/IP stack. A model of the dynamics of the TCP-NC protocol combined with random early detection (RED) as active queue management mechanism is considered to study the network equilibrium and stability properties. The existence and uniqueness of an equilibrium point is demonstrated and characterized in terms of average throughput, loss rate, and queue length. Global stability is proved in absence of forward delay, and the effects of the NC redundancy factor and of the delay on the local stability of TCP-NC-RED are studied around the equilibrium. The fairness of TCP-NC with respect to TCP-Reno-like protocols is also studied. A version of TCP-NC with adaptive redundancy factor (TCP-NCAR) is also introduced. Results provided by the proposed model are compared to those obtained by simulation for N sources sharing a single link. TCP-NC-RED becomes unstable when delay or capacity increases, as TCP-Reno does, but also when the redundancy factor increases. Its stability region is characterized as a function of the redundancy factor. If TCP-NC and TCP-Reno share the same links, TCP-NC is fair with TCP-Reno-like protocols when no redundancy is added. Simulations show that TCP-NCAR is able to compensate losses on the wireless parts of the network.

This paper considers a multicast scenario and compares the average reception quality obtained when combining multiple description coding (MDC) and network coding (NC). Plain (single description) network coding (NC-SDC) serves as reference. In the considered scenario, a single source is multicast to several receivers with various channel conditions. Contrary to a NC-SDC scheme, unable to recover the coded packets when not enough combinations of packets have been received, NC of MDC packets allows a more progressive quality improvement with the number of received packets, and a reduction of the effect of the quantization noise when MDC is performed via frame expansion before quantization. Considering a probability distribution for the bit transition probability during transmission to any user in the multicast group, the expected signal-to-noise ratio is evaluated. Performance comparisons are made for various error distributions, field sizes, and MDC methods (via frame expansion and correlating transform).