Unsupervised Learning for Secure Short-Packet Transmission Under Statistical QoS Constraints

We maximize the effective secrecy throughout of a wireless system where the access point transmits confidential short packets to an intended user in the presence of an eavesdropper. To find the optimal power control policy under statistical quality-of-service and average transmit power constraints, we formulate a constrained functional optimization problem which does not have closed-form solution. To address this, we propose an unsupervised learning algorithm to solve the problem, where a deep neural network (DNN) is used to approximate the power control policy. Then, we train the parameters of the DNN by a primal-dual method. To provide more insights and verify the effectiveness of unsupervised learning, we derive the closed-form solution in a special case. Using numerical results, we show that the learning-based power control policy rapidly approaches the closed-form solution in the special case and can satisfy the constraints in general cases.