A computer's image is sent to the monitor as a stream of RGB pixels, with their intensities encoded as voltage levels in the signal traveling through the cable. During transmission, these signals give of unwanted electromagnetic emissions, which can be picked up using TempestSDR software to reconstruct the original image. However, the resulting image is often noisy and hard to read. With standard software defined radios (SDR), you can only eavesdrop on the image from a short distance, about 50 cm. In this paper, we propose two methods to improve the quality of the captured image. First, we used a directional Yagi-Uda antenna, a signal amplifier, and a band-pass iflter. This setup extended the capture distance to around 20 meters. Even so, the image remained noisy, so we turned to deep learning for our second approach. We created scripts to automate TempestSDR and the dataset generation process, producing a large dataset covering typical everyday computer use. We selected two convolutional neural networks designed for image reconstruction, trained them on our dataset, and evaluated their performance using standard image quality metrics. The results showed a significant improvement in reconstructed image quality across all tested metrics. Finally we integrated the trained models into TempestSDR, making high-quality image reconstruction much easier for users. Our findings highlight potential vulnerability of display devices and emphasize the need for preventive measures to enhance their security.
Confidentiality and privacy are fundamental pillars of information security. While conventional approaches focus primarily on mechanisms like authentication, encryption, and access control, hardwarebased vulnerabilities receive considerably less attention. Among these, side-channel attacks exploit unintended information leakage arising from standard hardware operations.
Electromagnetic (EM) emissions have long been recognized as a source of interference in radio
communications. Historically, most eforts focused on minimizing such emissions to prevent signal
degradation. However, military and intelligence institutions also recognized a more critical implication:
the possibility of intercepting these emissions to reconstruct sensitive data. Because early research on
this topic remained classified [
Computer monitors (and display devices in general) present a particularly viable target in this
context. During image transmission from a computer to a monitor, unintentional EM emissions are
inadvertently produced. These emissions can be intercepted and analyzed to reconstruct the original
screen content. The first public demonstration of such an attack was conducted by Wim van Eck in
1985, who successfully recovered images from analogue CRT monitors [
Today, this type of attack can be executed using TempestSDR [
CEUR Workshop
ISSN1613-0073 efective range for a successful image reconstruction is limited to under 50 centimeters. Furthermore, the reconstructed images are typically significantly compromised by noise and frequently lack clarity.
This paper investigates methods for extending the efective attack range and improving the quality of reconstructed images. The following section outlines the principles of image transmission, its inherent vulnerabilities, and explains how TempestSDR processes EM emissions to recover image content. The subsequent sections present two methods for improving both the range and visual clarity of the reconstructed output.
Several video interfaces are used to transmit image data from a computer to a display device. These include both analogue and digital standards, each of which operates in a way that makes them susceptible to EM leakage.
VGA interface. Video Graphics Array (VGA) [
The video signal is transmitted as a continuous stream of pixels. These pixels form horizontal scan
lines, with each scan line ending in a short blanking interval marked by the HSync signal. Once all
scan lines for a frame have been transmitted, a vertical blanking follows, indicated by the VSync signal.
This process repeats continuously, producing a sequence of full-screen image frames.
HDMI interface. High-Definition Multimedia Interface (HDMI) [
Each pixel’s color components are represented as 8-bit values, which are converted into 10-bit
TMDS symbols. These symbols are then transmitted as a high-speed serial data stream. Like VGA, the
stream of pixels is organized into horizontal scan lines, which in turn form complete image frames.
Synchronization signals are embedded within the stream, analogous to HSync and VSync.
EM susceptibility. The core principle behind the attacks demonstrated by Van Eck and Kuhn is that
image transmission relies on rapidly changing voltages to represent pixel colour intensities [
When analyzing captured emissions from analogue standards, TempestSDR leverages the fact that image data is transmitted at a rate defined by the monitor’s pixel clock frequency , which is the product of display width , height , and frame rate . Then, the time required to transmit a single pixel is = 1/( ⋅ ⋅ ), (1) which results in impulses in the captured signal occurring at multiples of . In essence, the amplitude of each impulse corresponds to the intensity of the pixel being transmitted at that moment. As a result, the signal can be properly segmented according to pixel transmission periods, and the amplitude of each impulse can be used to infer the approximate grayscale value for the corresponding pixel.
For digital transmission, where each pixel intensity is represented using bits, the bit period is = 1/( ⋅ ⋅ ⋅ ). (2) TempestSDR reconstructs the image by averaging the bit-level signal over each pixel period. Note that for digital transmissions, in which Display Stream Compression (DSC) is used, e.g. HDMI 2.1, an attack with the described approach is not feasible. However, it may still be possible to eavesdrop on the computer screen itself, rather than the cable.
Beyond signal demodulation, TempestSDR also ofers automatic detection of the target monitor’s resolution and refresh rate, allowing relatively straightforward attacks even in cases where these parameters are initially unknown.
Baseline attack. An illustrative eavesdropping attempt involves a VGA monitor with a resolution of 1024×768 pixels operating at a 60 Hz refresh rate. Using a USRP x310 software-defined radio, the emissions were captured at a distance of approximately 50 centimeters. TempestSDR was tuned to the second harmonic frequency of the monitor’s pixel clock, which, in this case, is approximately 130 MHz. A close-up comparison of the original (reference) image and the reconstructed output obtained through this baseline setup is shown in Figure 1.
Although much attention is typically given to software-based processes such as signal demodulation, the ability to receive a clean and suficiently strong signal at a distance is equally important. In electromagnetic (EM) side-channel attacks, this makes hardware optimization a critical factor.
Without appropriate hardware, the efective capture range is limited to very short distances (approximately 50 centimeters in our baseline setup). To extend this range, we employ a high-gain antenna, a low-noise amplifier, and a band-pass filter. This setup not only increases the reception range, but also improves the signal-to-noise ratio, thereby enhancing the quality of the reconstructed images. Directional antennas. The most common types of directional antennas applicable for this scenario are the log-periodic antenna and the Yagi-Uda antenna. While both are directional, they difer in bandwidth and frequency gain. Log-periodic antennas typically cover a wider frequency range (e.g. from 100 MHz to 1 GHz), whereas Yagi-Uda antennas have a narrow bandwidth, around 3% from their centre frequency. On the other side, Yagi-Uda antennas commonly ofer a higher gain compared to log-periodic antennas.
Log-periodic antennas are more suitable when the frequency of interest is not known in advance,
making them ideal for exploratory attacks. In contrast, the Yagi-Uda antenna is optimal when the
monitor specifications are known, allowing the antenna to be tuned precisely to the required frequency
for maximum signal gain. Previous works also support this rationale: van Eck originally used a Yagi-Uda
antenna [
The antenna was connected to the SDR using a 50 Ω coaxial cable. Since the folded dipole has a higher characteristic impedance (typically from 200 Ω to 300 Ω), we implemented a 4:1 balun to match it to the SDR’s 50 Ω input. A TQP3M9037 low-noise amplifier was added in line, providing an additional 20 dB of gain. Using a Tektronix MDO4104C multi-domain oscilloscope, we observed substantial interference in the 90-120 MHz band, mostly from FM radio broadcasts. To protect the SDR’s input channel and to isolate the desired signal, we applied a band-pass filter tuned to 125-135 MHz. A schematic setup is shown in Figure 2.
Results. With inspiration from [
We tested the implemented system in a hallway with a metal structure, containing various reflective and absorptive objects. The target device, a laptop connected to a monitor via an HDMI-to-VGA converter, and a basic unshielded VGA cable, was positioned at one end of the hallway. The attacker remained in their room, while the Yagi-Uda antenna was placed in the hallway to capture emissions from the target. The test scenario is depicted in Figure 4, and reconstruction results from various distances are shown in Figure 5. Both folded and straight dipoles achieved similar results.
While using optimized hardware allowed us to capture clearer signals from longer distances, the reconstructed images often remained degraded due to noise and distortion. To address this issue, we explored the application of deep learning techniques, specifically convolutional neural networks (CNNs), for image post-processing and enhancement. Problem definition. The image restoration task can be formally formulated as recovering a clean image from a degraded observation , defined by the relationship = + , where represents degradation, e.g. noise.
In the context of electromagnetic side-channel attacks, previous works have demonstrated the
feasibility of using deep learning to restore screen content from captured emissions [
Our approach instead focuses on reconstructing entire frames obtained from actual TempestSDR
captures. A similar objective was pursued in the Deep-Tempest paper [
more suitable for practical, real-time attack scenarios.
traditional models that attempt to predict a clean image directly from a noisy input, DnCNN is trained to estimate the residual image, that is, the diference between the noisy and the clean image. This strategy simplifies the task by focusing the model on removing the noise component rather than reconstructing the entire image from scratch.
Let the noisy input image be = +
, where is the clean image and is additive noise. The model learns to approximate the residual mapping ℛ( ) ≈ , such that the clean image can be recovered via = − ℛ( )
. The network parameters Θ are optimized by minimizing the mean squared error (MSE)
between the desired residual images and the estimated ones from the noisy input:
1
2 =1
ℓ(Θ) =
∑ ∥ ℛ( ; Θ) − ( − ) ∥2,
(3)
where ( , )=1 denotes a set of noisy-clean image pairs [
The architecture uses a deep convolutional structure without pooling layers. DnCNN with the depth consists of: • Conv + ReLU (first layer) – 64 convolutional filters with the size of 3 × 3 × , where is the number of input channels, • Conv + BN + ReLU (hidden layers 2 to − 1 ) – 64 convolutional filters with the size of 3 × 3 × 64, with batch normalization between the convolution and activation function, • Conv (final layer) – filters with the size of 3 × 3 × 64.
We selected this network for its simplicity, relatively low parameter count (0.5 ⋅ 106 for = 17 layers), and its efectiveness.
image restoration tasks, including denoising, deblurring, and super-resolution.
DRUNet formulates the task as an optimization problem, where the goal is to obtain the clean image from the degraded image = () +
, where represents degradation independent of noise, and
is noise [
From a Bayesian perspective, DRUNet predicts the approximation ̂ of the clean image by solving the Maximum A Posteriori (MAP) problem, which is given as: =̂ arg max log ( |) + log (), where log ( |)
represents the log-probability of the observation of with the given , log () represents a prior probability of the clean image, independent of the observation . Formally, we can rewrite (4) as a minimization of the function: 1 2 2 (4) (5) =̂ arg min ∥ − () ∥
2 +ℛ(),
2 is the data term, which ensures that the solution respects degradation, while
where 1
ℛ()
2 2 ∥ − () ∥
is the regularization term, which enforces the desired properties on the solution [
DRUNet’s architecture is based on the U-Net model, which features an encoder-decoder structure
with skip connections. Each encoder stage consists of convolutional layers that downsample the input
and extract features. The decoder stages then upsample these features to reconstruct the original
image. DRUNet consists of a four-layer decoder (and encoder) with channel depths of 64, 128, 256,
and 512, respectively. Each stage includes four residual blocks with the ReLU activation function, and
downsampling/upsampling is performed via strided/transposed convolutions [
DRUNet normally expects both a degraded image and a noise level map as its input. Since the exact noise profile in our case is unknown, the noise map channel is set to zero.
We selected DRUNet for its strong denoising performance on various benchmark datasets [
To simulate realistic screen content, we manually collected 650 screenshots of common desktop environments, popular websites, and various graphical user interfaces. Since improving text legibility is a primary objective of this work, we extended the dataset with 200 synthetically generated images containing randomly rendered text. These were created using a custom Python script built on the Selenium framework, which randomly selects font styles and sizes, renders the text in a browser window, and captures the resulting screen.
We fully automated the data acquisition process in TempestSDR using the PyAutoGUI library. A custom script was developed to iteratively display each reference image on the target monitor and trigger a snapshot in TempestSDR. For every reference image, eight captures were performed, each from a diferent distance and antenna orientation. The captures include images reconstructed from VGA and HDMI cable, as well as from HDMI-VGA converter emissions. These variations simulate real-world conditions, as even small changes in antenna placement can significantly afect the output image.
The snapshots produced in TempestSDR contain not only the reconstructed image but also padding caused by synchronization signals present during image transmission. Since the selected CNNs operate on pixel-wise comparisons between reference and captured images, this padding had to be removed. To accomplish this, we implemented a script that uses template matching from the OpenCV library. If the script failed to locate the reference image within the snapshot, the snapshot was deemed invalid and excluded from the dataset. Otherwise, the padding was removed, and the resulting image was retained for training. resulting subsets.
Finally, the dataset was divided into training, testing, and validation subsets in a ratio of 80:10:10. The ratio of images captured from HDMI, VGA cable, and HDMI-VGA converter was preserved in the
PSNR Training. Both DnCNN and DRUNet were trained on the training subset with the goal of minimizing the MSE loss function. The initial learning rate was set to 10−4, and the Adam Optimizer was employed. The input image pairs were split into patches of size 256 × 256 pixels by the dataloader, and the batch size was set to 32. Each model was trained for 100 epochs, with checkpoints saved every 10 epochs. Results. After training the models, we evaluated the performance with standard image quality metrics: Peak Signal-to-Noise Ratio (PSNR), Mean Squared Error (MSE), and Structural Similarity Index Measure (SSIM). To better assess text readability, we additionally measured the Character Error Rate (CER), defined as the ratio of incorrectly recognised characters to the total number of characters in the reference image. We also recorded the average reconstruction time per image.
From the results in Table 1, it is evident that using DRUNet significantly enhances the quality of reconstructed images. The CER was reduced by 40%, greatly enhancing text readability in the images. While DnCNN also outperforms TempestSDR, it lags behind DRUNet across all image quality metrics. Its primary advantage is faster inference, requiring only half the reconstruction time compared to DRUNet. A real-world example of image reconstruction using both trained models is shown in Figure 6. Integration into TempestSDR. Currently, in order to perform an attack using the trained models, the user has to manually capture an image in TempestSDR, remove padding from the captured image, run a separate script for image enhancement using a trained model, and then save the improved result. To streamline and simplify the attack process, we integrated a model inference functionality directly into TempestSDR.
Two new buttons were added to the GUI: Select Model (DRUNet or DnCNN) and Load Model (to specify the path to the trained model). Once loaded, the user can utilize the Take snapshot and enhance option to automatically enhance the captured image using the selected model. While removing the padding present in TempestSDR snapshots, contour detection is used to locate the region corresponding to the screen content. The added functionality can be seen in Figure 7.
We demonstrated that electromagnetic side-channel attacks on monitors, previously limited by short range and poor image quality, can be greatly improved through non-expensive hardware and software enhancements. By designing a directional Yagi-Uda antenna and applying a band-pass filter and amplifier, we extended the efective capture distance from 50 cm to 20 m. To further improve image quality, we trained CNNs on real TempestSDR captures, achieving up to a 40% improvement in text recognition accuracy. Finally, we integrated the models into TempestSDR, enabling practical and repeatable attacks with one-click image restoration.
This publication is the result of support under the Operational Program Integrated Infrastructure for the project: Advancing University Capacity and Competence in Research, Development and Innovation (ACCORD, ITMS2014+:313021X329), co-financed by the European Regional Development Fund.
During the preparation of this work, the authors used GPT-4o for grammar and spell check. After using the tool, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. The sources for our modified TempestSDR, dataset, and trained models are available online: