Accurate peak determination from noise-corrupted photoplethysmogram (PPG) signal is the basis for further analysis of physiological quantities such as heart rate and heart rate variability. In the past decades, many methods have been proposed to provide reliable peak detection. These peak detection methods include rule-based algorithms, adaptive thresholds, and signal processing techniques. However, they are designed for noise-free PPG signals and are insufficient for PPG signals with low signal-to-noise ratio (SNR). This paper focuses on enhancing PPG noise-resiliency and proposes a robust peak detection algorithm for noise and motion artifact corrupted PPG signals. Our algorithm is based on Convolutional Neural Networks (CNN) with dilated convolutions. Using dilated convolutions provides a large receptive field, making our CNN model robust at time series processing. In this study, we use a dataset collected from wearable devices in health monitoring under free-living conditions. In addition, a data generator is developed for producing noisy PPG data used for training the network. The method performance is compared against other state-of-the-art methods and tested in SNRs ranging from 0 to 45 dB. Our method obtains better accuracy in all the SNRs, compared with the existing adaptive threshold and transform-based methods. The proposed method shows an overall precision, recall, and F1-score 80%, 80%, and 80% in all the SNR ranges. However, these figures for the other methods are below 78%, 77%, and 77%, respectively. The proposed method proves to be accurate for detecting PPG peaks even in the presence of noise.