In the digital age, the importance of privacy protection is increasingly evident. However, you might not have anticipated that even the electromagnetic radiation from HDMI cables could potentially serve as a channel for information leakage. Recently, a research team from the Engineering School of the University of the Republic in Uruguay has achieved a remarkable feat by recovering original image content from electromagnetic signals leaked through HDMI cables using AI technology.
The core of this research is an end-to-end AI model focused on text restoration, which reduces the character error rate of HDMI signals to about 30%. This might sound abstract, but imagine the content on your computer screen on the far right, and the result output by the AI model in the middle, and you'll grasp the震撼力 of this technology.
We know that compared to analog signals, digital signals like HDMI are more difficult to recover due to the 10-bit encoding, which increases bandwidth and results in a nonlinear mapping between the signal and pixel intensity. However, the emergence of this technology has made the once elusive electromagnetic waves decipherable.
The research team first captures the electromagnetic waves emitted by the HDMI cable and connectors using an antenna, then receives these signals through a Software Defined Radio (SDR) device, converting them into digital samples. Next, software tools are used to process the signals, extract image data, and finally input it into the AI model for image recognition and enhancement.
The key lies in the use of a Deep Residual UNet (DRUNet), an encoder-decoder structured convolutional neural network, particularly suited for image restoration tasks. By optimizing the network structure and training process, DRUNet significantly enhances the quality of image restoration, especially in terms of text readability.
To validate this technology, the team constructed a dataset containing about 3,500 samples for testing. The results show that on real datasets, the model using complex samples outperforms in multiple evaluation metrics. Traditional methods have a character error rate of over 90% on real datasets, while their model can reduce this to 35.3%.
This research not only demonstrates the potential of AI in the field of information security but also reminds us that even seemingly secure HDMI connections may pose a risk of information theft. However, the research team also proposes preventive measures, such as adding low-level noise to the monitor image or using background gradients, which can effectively reduce the success rate of electromagnetic leakage.
Project Repository: https://github.com/emidan19/deep-tempest
Paper Link: https://arxiv.org/pdf/2407.09717