Complex for automated determination of type of electronic devices for covert information removal and unauthorized operation of radio transmitting modules based on semi analysis
A system using inductor coils and convolutional neural networks to analyze PEMIN signals for automated recognition of electronic devices and modes addresses the challenge of identifying information leakage channels, enhancing data protection.
Patent Information
- Authority / Receiving Office
- RU · RU
- Patent Type
- Patents
- Current Assignee / Owner
- ГУСТОВ ВЛАДИМИР ВЛАДИМИРОВИЧ
- Filing Date
- 2025-03-03
- Publication Date
- 2026-07-07
Smart Images
Figure 00000001_ABST
Abstract
Description
[0001] The invention relates to the field of information protection and identification of information leakage channels.
[0002] The invention technology is based on receiving, processing, categorizing and analyzing unique components of PEMIN signals.
[0003] The technical result of the proposed product is automated recognition of the type and operating mode of an electronic or electromechanical device. Information leakage channels will be identified by a neural network detecting specific patterns based on their similarity to a compiled database of unique signals from PEMIN sources.
[0004] The essence of the invention lies in the recognition of sources of low-frequency electromagnetic fields.
[0005] In the proposed invention, inductor coils are used as antennas, which receive PEMIN signals, then they are transmitted to the input of the operational amplifier, after which they are converted into audio format.
[0006] The PEMIN signals are then transmitted to the personal computer's sound card, where they are presented as a microphone input to the software. The software then constructs a spectrogram and sonogram of the audio signals.
[0007] Following this, images of key components characterizing the type / operating mode of the selected electronic device are created from sonogram fragments. These components are categorized, creating a database that serves as a training dataset for convolutional neural networks. This process results in the recognition of PEMIN sources.
[0008] Thus, the software, in this case, allows for the automation of the process of recognizing sources of low-frequency electromagnetic fields and significantly simplifies data analysis for further processing and decision-making.
[0009] The structural design of the software of the proposed product should have the following form:
[0010] The audio signal is visualized in real time, with the construction of a spectrum (dependence of the signal level on time) and a sonogram.
[0011] Afterwards, the sonogram is saved as an image at specified intervals (3-5 s) and transmitted to the neural network for recognition.
[0012] The neural network and search target recognition are formed based on a database of signals of categorized behavior patterns and types of electronic and electromechanical devices.
[0013] When a match is found above the specified threshold, the software should automatically notify the operator and add a new entry to the event log.
[0014] In the context of convolutional neural networks, a database / dataset is a set of images on which the model is trained. The database is used for training and testing the model.
[0015] As the database accumulates, patterns emerge that allow for the identification of various electronic device modules and the monitoring of their operating modes. To facilitate sonogram comparison, they must be structured into groups. This is used to train a convolutional neural network, enabling highly accurate automated recognition of the electronic device type based on the PEMIN signal.
[0016] The circuit diagram, PCB layout and PCB prototype have been developed for the proposed product.
[0017] The main idea of using the proposed product is the automated detection of electronic and electromechanical devices, their operating modes by analyzing PEMIN signals.
[0018] The scope of application of the proposed product covers such areas as working with state secrets, preserving commercial secrets in a competitive environment, as well as preserving personal data, the need to protect which is enshrined at the legislative level by the Constitution of the Russian Federation, the Labor Code of the Russian Federation, and Federal Law No. 152-FZ of July 27, 2006 "On Personal Data".
[0019] Brief description of drawings
[0020] Fig. 1 shows the basic electrical circuit diagram of the proposed product.
[0021] Fig. 2 shows the printed circuit board topology (top view) of the proposed product.
[0022] Examples of patterns that emerge as data accumulates
[0023] The sample provided forms a training database for a convolutional neural network from measurements of 5 types of electronic / electromechanical devices, each of which creates a unique sonogram characterizing the type of electronic device and / or its operating mode (depending on the training database).
[0024] Types of electronic devices:
[0025] 1) Wireless chargers (Fig. 3 - Fig. 8)
[0026] 2) Pulse power supplies (Fig. 9 - Fig. 16)
[0027] 3) Microcontrollers (Fig. 17 - Fig. 24)
[0028] 4) Radio transmitters (GSM module) (Fig. 25 - Fig. 32)
[0029] 5) Electromechanical devices (Fig. 33 - Fig. 40)
[0030] When analyzing the provided sonograms (Fig. 3 - Fig. 40), it is possible to identify key patterns that allow us to identify the operating modes of various electronic devices, as well as determine their potential affiliation with devices for covert information collection or radio transmitting modules.
[0031] 1. Unique patterns in spectral characteristics
[0032] Every electronic device generates electromagnetic radiation (EMR), which is reflected in the form of certain frequency patterns on sonograms. Examples of wireless chargers (Fig. 3 - Fig. 8), pulse power supplies (Fig. 9 - Fig. 16), microcontrollers (Fig. 17 - Fig. 24), radio transmitters (Fig. 25 - Fig. 32) and electromechanical devices (Fig. 33 - Fig. 40) show that:
[0033] - The number of vertical lines and the intervals between them are strictly specific for each device and its operating mode.
[0034] - Devices in different operating modes exhibit changes in the intensity and location of lines, which creates a unique power spectral density graph.
[0035] 2. Examples of patterns
[0036] Wireless charger: Sonograms (Fig. 3 - Fig. 8) show persistent, high-intensity vertical lines, which are caused by inductive energy transfer processes. The frequency spectrum also shows characteristic noise associated with the operation of the charging management system.
[0037] Pulse Power Supplies: Sonograms (Fig. 9 - Fig. 16) show the generation of horizontal bands that correspond to the converter operating frequencies. The intensity and number of bands vary depending on the load.
[0038] Microcontrollers: The sonograms (Fig. 17 - Fig. 24) reveal complex regular structures reflecting the process of clocking and switching control signals in the microcontroller. These patterns change when executing different programs or when the clock frequency changes.
[0039] Radio Transmitters: The sonograms (Fig. 25 - Fig. 32) show periodic vertical lines, the distances between which depend on the current operating mode of the device (e.g., the start of data transmission). These lines represent the frequencies used by the module to transmit information.
[0040] Electromechanical devices: On the sonograms (Fig. 33 - Fig. 40) clearly defined harmonics are typical for motors and generators, since they operate at fixed frequencies.
[0041] The presence of mechanical vibrations - manifests itself in the form of repetitive periodic bursts or rhythmic stripes.
[0042] Combination of electrical and mechanical noise - the superposition of electromagnetic interference with vibration characteristics is visible.
[0043] Amplitude modulation - frequencies can change depending on the load on the engine or mechanical system.
[0044] Low frequency components - reflect the operation of gearboxes, electromagnetic brakes or stepper motors.
[0045] Flicker noise (1 / f noise) - often found in low-power electromechanical systems such as servos.
[0046] 3. Data Transformation for Analysis
[0047] The patterns revealed in sonograms form the basis for constructing a database of reference signals. This data is processed for:
[0048] - Comparisons with samples from the database.
[0049] - Identification of devices and determination of their operating mode.
[0050] - Detection of deviations that may indicate covert information collection.
[0051] 4. The Role of Convolutional Neural Networks
[0052] Sonograms (Fig. 3 - Fig. 40) are converted into fixed-size images (248 × 246 pixels), which are then used to train convolutional neural networks. Unique frequency patterns present in the sonograms allow the neural network to classify devices and their operating modes with high accuracy. The neural network is trained to detect:
[0053] - Frequency ranges and their distribution.
[0054] - Changes in signal intensity.
[0055] - Periodicity and intervals between signals.
[0056] Thus, the method enables the automation of device identification and the detection of unauthorized activity. The advantage of this approach is its versatility: each device type and its operating mode generates a unique frequency signature that is easily detected on sonograms and can be interpreted by a neural network.
[0057] Neural Network Training Algorithms
[0058] Preparing training sample:
[0059] Collecting PEMIN signals from various types of electronic and electromechanical devices in various operating modes. Each signal is recorded in audio format and divided into 3-second segments.
[0060] Sample code for splitting audio files
[0061] python
[0062] from pydub import AudioSegment
[0063] from pydub.utils import make_chunks
[0064] import os
[0065] from strgen import StringGenerator as SG
[0066] # Creating directories for storing images and split audio files
[0067] if os.path.exists('Images'):
[0068] os.remove('Images')
[0069] os.mkdir('Images')
[0070] else:
[0071] os.mkdir('Images')
[0072] if os.path.exists('Splitted'):
[0073] os.remove('Splitted')
[0074] os.mkdir('Splitted')
[0075] else:
[0076] os.mkdir('Splitted')
[0077] # Split audio files into 3-second chunks
[0078] for root, dirs, files in os.walk("Audios", topdown=False):
[0079] for name in dirs:
[0080] if not os.path.exists(fSplitted / {name}'):
[0081] os.mkdir('Splitted / ' + name)
[0082] path = os.path.join(root, name)
[0083] for root_2, dirs_2, files_2 in os.walk(path, topdown=False):
[0084] for file in files_2:
[0085] src = path + " / "+file
[0086] myaudio = AudioSegment.from_file(src, src.split(".")[-1])
[0087] chunk_length_ms = 3000 # pydub calculates in millisec
[0088] chunks = make_chunks(myaudio, chunk_length_ms) # Make chunks of 3 seconds
[0089] # Export all individual fragments as WAV files
[0090] for i, chunk in enumerate(chunks):
[0091] if i == len(chunks) - 1:
[0092] continue
[0093] chunk_name = "Splitted / " + name + " / " + SG(r"[\w]{30}").render() + ".wav"
[0094] print("exporting", chunk_name)
[0095] chunk.export(chunk_name, format="wav")
[0096] For each audio fragment, a power spectral density graph is generated using the Fourier transform. Spectrograms are saved as 248x246 pixel images.
[0097] Formation of training database:
[0098] Spectrograms are classified by device type and their operating modes. The database contains categories such as:
[0099] Wireless chargers
[0100] Electromechanical devices
[0101] Microcontroller tools
[0102] Pulse power supplies
[0103] Radio transmitters
[0104] For each device, its operating modes are recorded.
[0105] Neural network architecture:
[0106] The network is based on a convolutional neural network (CNN) based on the pre-trained VGG16 architecture, adapted for spectrogram analysis. The upper layers of the network are supplemented with:
[0107] Global Average Pooling 2D
[0108] A fully connected layer of 1024 neurons with a ReLU activation function
[0109] Output layer with the number of neurons equal to the number of device classes and softmax activation function.
[0110] Learning process:
[0111] The model is trained on spectrograms using the backpropagation method and the Adam optimizer. The loss function is categorical_crossentropy.
[0112] During the training process, the data is divided into training (80%) and validation (20%) samples.
[0113] Optimization and testing:
[0114] After training, the model is tested on new data to evaluate the accuracy of recognizing device types and operating modes.
[0115] The model is converted to ONNX format for integration with the hardware and software system.
[0116] Automated detection of information leakage channels:
[0117] The output data of the neural network is analyzed to identify characteristic features of the operation of radio transmitting modules, microcontrollers and other potentially dangerous devices.
[0118] The classification results are used to make decisions about the detection and classification of information leakage channels.
Claims
1. A system for automated determination of the type of electronic devices for covert information collection and unauthorized operation of radio transmitting modules through the analysis of side electromagnetic emissions and interference (SEMI), characterized by the fact that it contains an antenna made in the form of an inductance coil, capable of receiving SEMI signals, an operational amplifier capable of converting the received signals into an audio signal, and a computing device capable of: receiving an audio signal through a sound card; construction of a spectrogram and sonogram of an audio signal; formation of images of key components of signals characterizing the type and / or operating mode of an electronic or electromechanical device from fragments of a sonogram; categorization of the specified images and formation of a database; using the specified database as a training array for a convolutional neural network; recognition of the type and / or operating mode of electronic and electromechanical devices based on the analysis of sonograms using a convolutional neural network.
2. The complex according to paragraph 1, additionally including a module for generating spectral power densities, configured to process an audio signal in sonogram format and generate graphs of the spectral power density in a resolution of 248×246 pixels for subsequent analysis.