In this study, we propose a differentiable layer for OFDM-based autoencoders (OFDM-AEs) to avoid high instantaneous power without regularizing the cost function used during the training. The proposed approach relies on the manipulation of the parameters of a set of functions that yield complementary sequences (CSs) through a deep neural network (DNN). We guarantee the peak-to-average-power ratio (PAPR) of each OFDM-AE symbol to be less than or equal to 3 dB. We also show how to normalize the mean power by using the functions in addition to PAPR. The introduced layer admits auxiliary parameters that allow one to control the amplitude and phase deviations in the frequency domain. Numerical results show that DNNs at the transmitter and receiver can achieve reliable communications under this protection layer at the expense of complexity.