Advances in software defined radio (SDR) technology allow unprecedented control on the entire processing chain, allowing modification of each functional block as well as sampling the changes in the input waveform. This article describes a method for uniquely identifying a specific radio among nominally similar devices using a combination of SDR sensing capability and machine learning (ML) techniques. The key benefit of this approach is that ML operates on raw I/Q samples and distinguishes devices using only the transmitter hardware-induced signal modifications that serve as a unique signature for a particular device. No higher-level decoding, feature engineering, or protocol knowledge is needed, further mitigating challenges of ID spoofing and coexistence of multiple protocols in a shared spectrum. The contributions of the article are as follows: (i) The operational blocks in a typical wireless communications processing chain are modified in a simulation study to demonstrate RF impairments, which we exploit. (ii) Using an overthe- air dataset compiled from an experimental testbed of SDRs, an optimized deep convolutional neural network architecture is proposed, and results are quantitatively compared with alternate techniques such as support vector machines and logistic regression. (iii) Research challenges for increasing the robustness of the approach, as well as the parallel processing needs for efficient training, are described. Our work demonstrates up to 90-99 percent experimental accuracy at transmitter- receiver distances varying between 2-50 ft over a noisy, multi-path wireless channel.