Waveform Classification from TEMPEST Attacks

The paper addresses the classification of typical baseband signals to be found in computing devices picked up without any physical contact, just using RF probing. The automatic classification applies a convolutional neural network and alternatively, by tracing dedicated signal properties that are specific for certain signal formats. The neural network classification is based on a time-frequency transform preprocessing. We selected a low-resolution Short-Time Fourier Transform (STFT), but other options, such as the Wigner-Ville Distribution (WVD) and the Continuous Wavelet Transform (CWT) outputs are also presented. We analyzed simulated and measured signals. We found that classification is possible with very high reliability, even based on low-resolution representations that the human eye cannot distinguish, such as for Ethernet 100-BaseT and 1000-BaseT. Both studied approaches can be nicely combined for enhanced classification performance. We briefly discuss jamming as a possible countermeasure to protect against eavesdropping.