Any device and system currently manufactured in the world needs to comply with EMC (electromagnetic compatibility) standards aimed at ensuring a correct performance in an inevitably electromagnetically polluted environments. One of the most important is the case of radiated tests, where external EM (electromagnetic) waves impinging on the SUT (system under test): in practice, it is of interest for a laptop exposed to a nearby mobile phone or an airplane landing in proximity of a radar station. The level of immunity of an SUT is typically assessed in facilities where one or more impinging waves are generated and the subsequent impact on the SUT is assessed. Any type of test currently used is based on a one-to-one interpretation, where each test configuration (e.g., direction of arrival of the wave) gives rise to a measurable effect. Although delivering a simple interpretation of the results, this approach has strong limitations: practical constraints (time, costs) impose a maximum number of test configurations that cannot in any way be regarded as exhaustive of all possible threats; single test configurations are separately defined and the corresponding results are analyzed just looking for the worst-case among the few configurations tested. A more effective and robust approach would rather consist in devising a finite number of tests in such a way as to explore the general behavior of the SUT, to any kind of external interferer. In other words, it is possible, as done in many other fields, to apply a learning approach, where the outcome is a better understanding of the SUT behavior. The availability of a behavioral model of the SUT will allow predicting its behavior to any external threat, without requiring further tests. Two direct results are therefore possible: 1) to reduce the number of tests, thus speeding up the time spent in a test facility (cost reduction); 2) the certainty of knowing beforehand what could be the worst-case configurations for an external interferer (risk management, accuracy). Current test approaches are incapable of delivering these advantages, since the tests are intended to probe single responses to single interferer scenarios rather than a global behavior. This project aims at defining the tools needed for this change of paradigm, by exploiting the remarkable properties of the TREC (time-reversal electromagnetic chamber), a new test facility recently developed in Supelec. The TREC has been demonstrated to allow a flexible generation of wavefronts, under real-time conditions, without recurring to expensive arrays of sources, but rather taking advantage of diffusive media. Any combination of the parameters of the interferer can be generated, even multipath scenarios, by means of signal-processing techniques. This project will allow extracting a macromodel of the SUT, relating a parametric description of any interferer scenario, to the inner state variables of the SUT: this can be represented as a generalized transfer function of the system, though it will also be able of reproducing the likely non-linear behavior of the SUT electronics. The main tools used for this feature will come from statistical inference, widely used in many diverse fields involving complex interactions, ranging from sociology to data-mining. In practice, the proposed approach will allow a direct identification of the position and nature of coupling paths (e.g., faults in the SUT shielding), a precious tool in R&D phases. Moreover, the non-linear macromodelling of the SUT behavior will provide direct information about the most critical interferer scenario that could put an SUT in fault: this is of interest for industrial as well as defense issues, e.g., in aeronautics and electronic warfare. The result of this project will be the definition of a new research field, EMC imaging, which will pave the way to the integration to EMC testing of advanced signal-processing techniques, already successfully applied in fields such as radar imaging.