Noisy-Intermediate-Scale-Quantum (NISQ) devices are nowadays starting to become available to the final user, hence potentially allowing to show the quantum speedups predicted by the quantum information theory. However, before implementing any quantum algorithm, it is crucial to have at least a partial or possibly full knowledge on the type and amount of noise affecting the quantum machine. Here, by generalizing quantum generative adversarial learning from quantum states (Q-GANs) to quantum operations/superoperators/channels (here named as SuperQGANs), we propose a very promising framework to characterize noise in a realistic quantum device, even in the case of spatially and temporally correlated noise (memory channels) affecting quantum circuits. The key idea is to learn about the noise by mimicking it in a way that one cannot distinguish between the real (to be sensed) and the fake (generated) one. We find that, when applied to the benchmarking case of Pauli channels, the SuperQGAN protocol is able to learn the associated error rates even in the case of spatially and temporally correlated noise. Moreover, we also show how to employ it for quantum metrology applications. We believe our SuperQGANs pave the way for new hybrid quantum-classical machine learning protocols for a better characterization and control of the current and future unavoidably noisy quantum devices.