Deep-tempest: Using Deep Learning to Eavesdrop on HDMI from its Unintended Electromagnetic Emanations

Summary

In this project we have extended the original gr-tempest (a.k.a. Van Eck Phreaking or simply TEMPEST; i.e. spying on a video display from its unintended electromagnetic emanations) by using deep learning to improve the quality of the spied images. See an illustrative diagram above. Some examples of the resulting inference of our system and the original unmodified version of gr-tempest below.

Video demo

We are particularly interested in recovering the text present in the display, and we improve the Character Error Rate from 90% in the unmodified gr-tempest, to less than 30% using our module. Watch a video of the full system in operation:

How does it works? (and how to cite our work or data)

You can find a detailed technical explanation of how deep-tempest works in the reference below. If you found our work or data useful for your research, please consider citing it as follows:

@misc{fernández2024deeptempestusingdeeplearning,
      title={Deep-TEMPEST: Using Deep Learning to Eavesdrop on HDMI from its Unintended Electromagnetic Emanations}, 
      author={Santiago Fernández and Emilio Martínez and Gabriel Varela and Pablo Musé and Federico Larroca},
      year={2024},
      eprint={2407.09717},
      archivePrefix={arXiv},
      primaryClass={cs.CR},
      url={https://arxiv.org/abs/2407.09717},
      note={Submitted}
}

Data

In addition to the source code, we are also open sourcing the whole dataset we used. Follow this dropbox link to download a ZIP file (~7GB). After unzipping, you will find synthetic and real captured images used for experiments, training, and evaluation during the work. These images consists of 1600x900 resolution with the SDR's center frequency at the third pixel-rate harmonic (324 MHz).

The structure of the directories containing the data is different for synthetic data compared to captured data:

Synthetic data

• ground-truth (directory with reference/monitor view images)

  • image1.png
  • ...
  • imageN.png

• simulations (directory with synthetic degradation/capture images)

  • image1_synthetic.png
  • ...
  • imageN_synthetic.png

Real data

• image1.png (image1 ground-truth)
• ...
• imageN.png (imageN ground-truth)

• Image 1 (directory with captures of image1.png)

  • capture1_image1.png
  • ...
  • captureM_image1.png

• ...

• Image N (directory with captures of image1.png)

  • capture1_imageN.png
  • ...
  • captureM_imageN.png

Code and Requirements

Clone the repository:

git clone https://github.com/emidan19/deep-tempest.git

Both gr-tempest and end-to-end folders contains a guide on how to execute the corresponding files for image capturing, inference and train the deep learning architecture based on DRUNet from KAIR image restoration repository.

The code is written in Python version 3.10, using Anaconda environments. To replicate the working environment, create a new one with the libraries listed in requirements.txt:

conda create --name deeptempest --file requirements.txt

Activate it with:

conda activate deeptempest

Regarding installations with GNU Radio, it is necessary to use the gr-tempest version in this repository (which contains a modified version of the original gr-tempest). After this, run the following grc files flowgraphs to activate the hierblocks:

• binary_serializer.grc
• FFT_autocorrelate.grc
• FFT_crosscorrelate.grc
• Keep_1_in_N_frames.grc

Finally run the flowgraph deep-tempest_example.grc to capture the monitor images and be able to recover them with better quality using the Save Capture block.

IIE Instituto de Ingeniería Eléctrica, Facultad de Ingeniería, Universidad de la República, Montevideo, Uruguay, http://iie.fing.edu.uy/investigacion/grupos/artes/

Please refer to the LICENSE file for contact information and further credits.