The electromagnetic radiation from HDMI cables can leak some display signals into the surrounding air. This wasn't a big issue before, but combined with AI, it's now possible to reverse-engineer the original image content.
A team from the Faculty of Engineering at the University of the Republic of Uruguay proposed an ### end-to-end model focusing on text recovery, which can reduce the character error rate (CER) of leaked signals (such as HDMI) to around 30%.
It's worth noting that compared to analog signals (like VGA), digital signals (such as HDMI) are more difficult to recover due to the increased bandwidth and non-linear mapping between signal and pixel intensity caused by 10-bit encoding.
At this level, the original content can basically be decrypted.
For a more intuitive understanding, let's look at one of the attack methods demonstrated by the team.
As a simple hint, the team ### used an antenna to intercept HDMI electromagnetic signals, and then used AI to attempt to "reconstruct" the original data.
Using AI Models
How exactly did they do it? The related research paper has been published on arXiv.
First, the team used ### antennas to capture electromagnetic waves emitted by HDMI cables and connectors.
Next, they used SDR (Software Defined Radio) devices to receive these electromagnetic signals and convert them into digital samples, which contain information from the original video signal but may also include noise and distortion.
Then, using software tools like ### gr-temest, they further processed the signals captured by SDR to extract image data.
This step may include operations such as filtering and sampling rate adjustment, with the aim of restoring the original form of the image as much as possible.
Finally, the processed signals were input into an ### AI model that can recognize and enhance key features in the image, thereby improving the clarity and readability of the image.
In summary, the entire process includes capturing electromagnetic signals, processing them using open-source software, and further processing using an end-to-end model.
As we can see, the ### key improvement in this research lies in the use of deep learning technology at the end.
The team used ### Deep Residual UNet (DRUNet), which is a convolutional neural network with an encoder-decoder structure suitable for image restoration tasks.
By optimizing the network structure and training process, DRUNet can significantly improve the quality of image restoration, especially in terms of text readability.
Error Rate Reduced by About 60%
So, how does this end-to-end model perform specifically?
To test it, they built a ### dataset containing about 3,500 samples, of which about 1,300 were real captured signals, and the rest were simulated signals.
Real samples were obtained through experimental setups, while simulated samples were generated using a GNU Radio simulator based on analytical models. These samples were used to train and evaluate the model.
The research shows that on the real dataset, the Pure Model using complex samples demonstrated ### the best performance across all evaluation metrics (PSNR, SSIM, CER).
Specifically, the traditional gr-tempest method using original image amplitudes had a CER ### exceeding 90% on the real dataset, while the Pure Model (using complex samples) reduced the CER to ### 35.3%.
At the same time, models trained on synthetic data may encounter performance degradation on real data.
However, ### through model fine-tuning, even using only ### 10% of real samples can achieve performance close to that of the Pure Model trained on all real samples.
To verify robustness, the model was tested with different sampling rates and display resolutions, and the results showed that ### some configuration changes might lead to significant performance degradation.
Although the team greatly improved the HDMI "cracking rate" with the new model, they also proposed ### corresponding countermeasures to prevent risks.
By ### adding low-level noise to the display image or using background gradients, the success rate can be effectively reduced.
Currently, the related research and dataset have been open-sourced. If interested, you can further read the paper.
Reference links: