Artificial intelligence edge calculation method for unmanned aerial vehicle and Internet of Things terminal

Abstract

The invention relates to the technical field of artificial intelligence, and discloses an artificial intelligence edge calculation method for an unmanned aerial vehicle and an Internet of Things terminal. The method comprises the following steps: performing asymmetric decoupling coding on original sensing data at an edge terminal, and generating task related representation and privacy sensitive representation; the privacy representation is destroyed, and only the task representation subjected to confusion compression is uploaded to the cloud; and after the cloud receives and recovers the representation, the task specialization reasoning unit completes high-precision AI reasoning. The system comprises an asymmetric decoupling coding unit deployed at a terminal and a task specialization reasoning unit at a cloud end. According to the method, through representation decoupling and cloud collaboration, the reasoning precision is remarkably improved while data privacy is guaranteed, and the terminal communication and calculation overhead is greatly reduced.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to an artificial intelligence edge computing method for drones and IoT terminals. Background Technology

[0002] With the deep integration of artificial intelligence and edge computing technologies, drones and IoT terminals are gradually taking on key roles in real-time perception, decision-making, and response in scenarios such as smart cities, emergency rescue, and agricultural monitoring. Edge computing effectively reduces cloud transmission latency and improves system response speed and data processing efficiency by offloading computing tasks to network edge nodes.

[0003] However, deploying high-precision AI models on resource-constrained edge devices while meeting data privacy compliance requirements has become a core challenge in current technological development. Especially in drone inspections or IoT monitoring tasks involving sensitive information (such as geographic location, personnel activity trajectories, and environmental images), using raw data directly for model inference without protection can easily lead to privacy breaches.

[0004] AI edge computing methods for drones and IoT terminals aim to achieve a synergistic optimization of efficient localized inference and privacy protection. This approach typically relies on mechanisms such as differential privacy, data anonymization, or federated learning to introduce perturbations before data upload or model training to hide individual information. However, existing technologies generally employ a globally uniform privacy noise injection strategy, applying a fixed intensity of perturbation to all input data to meet a pre-defined privacy budget constraint. While such rigid designs provide theoretical privacy guarantees, they neglect the differentiated requirements for inference accuracy across various tasks.

[0005] There is a fundamental contradiction between existing technologies and the accuracy of AI inference: On the one hand, while high-intensity perturbations can enhance privacy protection, they can also severely distort key features, leading to a significant decrease in the accuracy of highly sensitive tasks such as target detection and path planning. On the other hand, reducing the intensity of disturbances in order to maintain accuracy may lead to excessive exposure of sensitive information in non-critical tasks, violating the principle of minimum necessity.

[0006] Especially in dynamic task scenarios, such as when drones are searching for missing persons and need to accurately identify facial features, while only a blurred outline is needed for routine area patrols, the globally fixed noise mechanism cannot adaptively adjust according to the semantic importance of the task, resulting in a double waste of privacy overhead and accuracy loss. Therefore, there is an urgent need for an edge computing method that can dynamically adjust the intensity of privacy perturbation based on the criticality of the task, so as to maximize the performance of artificial intelligence inference while ensuring privacy compliance. Summary of the Invention

[0007] The purpose of this invention is to provide an artificial intelligence edge computing method for drones and IoT terminals, in order to solve the inherent contradiction between data privacy protection and model inference accuracy when edge devices perform artificial intelligence tasks, as well as the technical problems of huge data transmission bandwidth and computing resource consumption in the prior art.

[0008] To achieve the above objectives, this invention provides an AI edge computing method for drones and IoT terminals. This method constructs an asymmetric decoupled encoder at the edge terminal to decompose the original high-dimensional sensor data into two independent low-dimensional latent space vectors: task-related representations and privacy-sensitive representations. Only the task-related representations, after specific obfuscation transformations, are transmitted to a cloud server. The matching task-specialized inference unit on the cloud server then performs the final accurate analysis and decision-making. This method completely removes and discards privacy data at the source while retaining all semantic information necessary to complete the specified task. Thus, while ensuring absolute data privacy, it leverages the powerful computing capabilities of the cloud to achieve high-precision AI inference and significantly reduces the data transmission load on the edge terminal.

[0009] According to an aspect of the present invention, an artificial intelligence edge computing method for drones and Internet of Things (IoT) terminals is provided, comprising the following steps: On the edge computing terminal side, raw sensor data is collected in real time through the data acquisition unit; The asymmetric decoupling coding unit built into the edge computing terminal processes the raw sensing data to generate mutually orthogonal task-related potential representations and privacy-sensitive potential representations. The task representation obfuscation transmission unit of the edge computing terminal performs irreversible obfuscation transformation and compression encoding on the task-related potential representation to generate an obfuscated task data packet to be transmitted, and sends the obfuscated task data packet to the cloud computing server through the wireless communication module; the privacy-sensitive potential representation is destroyed immediately after generation and is not stored or transmitted in any form. The task representation receiving unit of the cloud computing server receives and parses the obfuscated task data packet to recover the task-related potential representation. The task-specific inference unit of the cloud computing server receives the recovered task-related potential representations, performs a preset artificial intelligence inference task based on them, and outputs the final inference result.

[0010] Furthermore, the data acquisition unit includes at least one sensor configured on a drone or IoT terminal, wherein the sensor is one or more of an image sensor, a lidar sensor, an infrared sensor, an inertial measurement unit, or a global positioning system module. The raw sensing data is a high-dimensional data stream containing timestamps, and its data format is video stream, point cloud data, thermal imaging matrix, or multi-axis time-series signal.

[0011] In one embodiment of the present invention, the asymmetric decoupling coding unit is a deep neural network model, the structure of which includes a shared convolutional feature extraction subunit, and parallel task-related representation generation subunits and privacy-sensitive representation generation subunits connected to the back end of the convolutional feature extraction subunit. The convolutional feature extraction subunit adopts a residual network structure, performing multi-layer convolution and pooling operations on the input raw sensor data to extract high-dimensional shared feature maps.

[0012] Both the task-related representation generation subunit and the privacy-sensitive representation generation subunit are composed of multi-layer fully connected networks, which respectively map the shared feature map into task-related latent representation vectors and privacy-sensitive latent representation vectors of predetermined dimensions.

[0013] Specifically, the training process of the asymmetric decoupling coding unit employs an adversarial training mechanism. This training process involves a generator, namely the asymmetric decoupling coding unit, and three independent discriminator networks: a task discriminator, a privacy discriminator, and a reconstruction discriminator. The overall training objective of the asymmetric decoupling coding unit is defined by a composite loss function, mathematically expressed as: Total Loss = Reconstruction Loss + First Weighted Coefficient Multiplied by Task Loss - Second Weighted Coefficient Multiplied by Privacy Adversarial Loss.

[0014] The reconstruction loss measures the difference between the reconstructed data using the task-related latent representations and the privacy-sensitive latent representations and the original sensor data; it is calculated as mean squared error. The task loss, generated by the task discriminator, measures the accuracy of completing a specific AI task using the task-related latent representations; it is calculated as cross-entropy loss.

[0015] The privacy-adversarial loss, generated by the privacy discriminator, measures the success rate of recovering privacy information from the task-related latent representations. During training, the asymmetric decoupling encoding unit strives to minimize the reconstruction loss and task loss while maximizing the privacy-adversarial loss, thereby forcing the encoder to expel all task-irrelevant privacy information into the privacy-sensitive latent representations, ensuring the semantic purity and privacy security of the task-related latent representations. The first and second weighting coefficients are pre-set scalar values ​​used to balance the contribution of each loss to the total loss.

[0016] Furthermore, the task representation obfuscation transmission unit includes a linear projection layer, a scalar quantization layer, and a data encapsulation module. The linear projection layer contains a fixed, non-square projection matrix that linearly maps the task-related potential representations from a first-dimensional space to a lower-dimensional second-dimensional space, achieving initial data compression and information obfuscation. The scalar quantization layer converts each floating-point value in the vector after linear projection into a fixed-point integer value according to a preset quantization bit width. The data encapsulation module combines the quantized integer sequence, along with a frame header containing a device identifier, timestamp, and data checksum, into a binary-formatted obfuscated task data packet.

[0017] In one embodiment of the present invention, the task-specialized inference unit of the cloud computing server is a deep neural network decoder trained in pairs with the asymmetric decoupled encoding unit. The input of the task-specialized inference unit is the recovered task-related latent representation obtained from the task representation receiving unit. The internal structure of the task-specialized inference unit is set according to the specific artificial intelligence inference task.

[0018] In the object detection task, the task-specialized inference unit employs a multi-layer deconvolutional network structure based on an attention mechanism. It progressively upsamples and decodes the input low-dimensional latent representation into an output tensor containing the object category, confidence score, and bounding box coordinates. In the semantic segmentation task, the task-specialized inference unit uses an encoder-decoder structure. Through a series of upsampling and skip connection operations, it restores the latent representation to a pixel-level segmentation mask with the same resolution as the original sensor data.

[0019] According to another aspect of the present invention, an artificial intelligence edge computing system for drones and Internet of Things (IoT) terminals is provided, comprising: An edge computing terminal, which is deployed on a drone or an Internet of Things (IoT) device; A cloud computing server that communicates with the edge computing terminal via a wireless network.

[0020] The edge computing terminal includes: The data acquisition unit is used to collect raw sensor data in real time; An asymmetric decoupling coding unit is used to process the original sensing data, generate mutually orthogonal task-related potential representations and privacy-sensitive potential representations, and destroy the privacy-sensitive potential representations immediately after generation. The task representation obfuscation transmission unit is used to perform irreversible obfuscation transformation and compression encoding on the potential representations related to the task, generate obfuscated task data packets, and send the data packets to the cloud computing server through the wireless communication module.

[0021] The cloud computing server includes: The task representation receiving unit is used to receive and parse the obfuscated task data packet to recover the task-related potential representation. The task-specific reasoning unit is used to receive the recovered task-related potential representations, and based on these, execute a preset artificial intelligence reasoning task and output the final reasoning result.

[0022] Furthermore, the asymmetric decoupling coding unit is implemented by a hardware accelerator, which is a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC). The neural network weight parameters of the asymmetric decoupling coding unit are stored in the read-only memory (ROM) of the hardware accelerator. The processor of the edge computing terminal is responsible for scheduling the data flow between the data acquisition unit and the asymmetric decoupling coding unit, and feeding the raw sensor data to the hardware accelerator through a direct memory access channel to achieve low-latency processing.

[0023] In one embodiment of the present invention, the convolutional feature extraction subunit of the asymmetric decoupling coding unit consists of a series of cascaded convolutional layers, batch normalization layers, and activation function layers, wherein the activation function is a modified linear unit. The final layers of the task-related representation generation subunit and the privacy-sensitive representation generation subunit do not use any activation function to output a linear latent space distribution. The dimension of the task-related latent representation is set to 256, and the dimension of the privacy-sensitive latent representation is set to 64.

[0024] Furthermore, the projection matrix used by the linear projection layer in the task representation obfuscation transmission unit is a random Gaussian matrix. This matrix is ​​generated and fixed during the system initialization phase and simultaneously distributed to all edge computing terminals and cloud computing servers as a shared key across the system. The quantization bit width of the scalar quantization layer is set to eight bits, mapping each floating-point value to an integer range of 0-255.

[0025] Specifically, after receiving the obfuscated task data packet, the task representation receiving unit of the cloud computing server first performs a data checksum verification to confirm the integrity of the data packet. Subsequently, it extracts the quantized integer sequence and restores it to a floating-point vector through an inverse quantization process.

[0026] Finally, it applies the generalized inverse matrix of the projection matrix to reverse map the floating-point vector from the second-dimensional space back to the first-dimensional space, thereby recovering a vector that is numerically similar to the task-related potential representation generated by the edge computing terminal, and then transmits it to the task-specialized inference unit.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention achieves decoupling between privacy protection and inference accuracy. By performing asymmetric decoupling encoding at the data source, it completely separates privacy information from task information at the representation level. By transmitting only task information, it fundamentally eliminates the risk of privacy data leakage. Unlike differential privacy technology, it does not sacrifice data utility, and it also avoids the huge computational overhead brought by homomorphic encryption technology.

[0028] 2. It ensures high-precision artificial intelligence reasoning capabilities. Since the core computation of the reasoning task is completed by a cloud server with unlimited resources, this invention can deploy and run extremely complex and large-scale task-specific reasoning models. Its accuracy far exceeds that of any lightweight model that can only run on an edge terminal, thus ensuring task performance in high-demand scenarios such as autonomous driving and security monitoring.

[0029] 3. Significantly reduces resource consumption at the edge terminal. This invention only requires a one-time, lightweight forward propagation encoding process at the edge terminal, with computational complexity far lower than performing a complete inference task. Simultaneously, by transmitting low-dimensional latent representation vectors instead of high-resolution raw sensor data, the amount of data transmitted is reduced by several orders of magnitude, significantly saving wireless communication bandwidth, reducing energy consumption of devices such as drones, and extending their operational endurance.

[0030] 4. Improved system robustness and security. The obfuscation transformation process adds an extra layer of security to data transmission. Even if the transmission channel is eavesdropped on, the intercepted data packets cannot be cracked or exploited by third parties due to their indirect interpretability and irreversibility. By employing adversarial mechanisms during the training phase, the system ensures the inherent immunity of the task representation to privacy attacks. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall technical solution architecture of an artificial intelligence edge computing method for drones and IoT terminals proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the asymmetric decoupling coding unit in this invention; Figure 3 This is a logical flowchart of data acquisition, decoupling encoding, and obfuscated transmission on the edge computing terminal side in this invention; Figure 4 This is a logical flow diagram of task representation reception and specialized reasoning on the cloud computing server side in this invention; Figure 5 This is a schematic diagram of the principle framework of the adversarial training mechanism and composite loss function of the asymmetric decoupled coding unit in this invention. Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the edge computing terminal and the cloud computing server in this invention. Detailed Implementation

[0032] Please refer to Figures 1 to 6 This invention provides an AI edge computing method for drones and IoT terminals. Its core lies in constructing an asymmetric decoupled coding structure at the edge to separate raw sensor data into task-related potential representations and privacy-sensitive potential representations. Only the obfuscated task-related representations are transmitted to the cloud server, where high-precision inference is performed. This method strictly follows the S-step numbering system to ensure the logical rigor of the technical solution and the clarity of the patent protection scope.

[0033] The method includes the following steps: S1, on the edge computing terminal side, raw sensor data is collected in real time through the data acquisition unit; S2, the asymmetric decoupling coding unit built into the edge computing terminal processes the original sensing data to generate mutually orthogonal task-related potential representations and privacy-sensitive potential representations; S3, the task representation obfuscation transmission unit of the edge computing terminal performs irreversible obfuscation transformation and compression encoding on the task-related potential representation to generate an obfuscated task data packet to be transmitted, and sends the obfuscated task data packet to the cloud computing server through the wireless communication module; the privacy-sensitive potential representation is destroyed immediately after generation and is not stored or transmitted in any form. S4, the task representation receiving unit of the cloud computing server receives and parses the obfuscated task data packet to recover the task-related potential representation; S5, the task specialization reasoning unit of the cloud computing server receives the recovered task-related potential representation and executes a preset artificial intelligence reasoning task based on it, and outputs the final reasoning result.

[0034] In step S1, the data acquisition unit is deployed on the drone body or an IoT terminal device, and its components include at least one sensor module. This sensor module can be one or more combinations of an image sensor, a lidar sensor, an infrared sensor, an inertial measurement unit, or a global positioning system module. The image sensor is used to capture a two-dimensional pixel array in the visible light band, outputting a red, green, and blue three-channel video stream at 30 frames per second, with each frame having a resolution of 1920×1080 pixels.

[0035] The lidar sensor emits pulsed laser beams and receives reflected echoes, generating three-dimensional point cloud data with a density of 100,000 points per second. Each point contains three-dimensional coordinates and a reflection intensity value. The infrared sensor outputs a thermal imaging matrix with dimensions of 640×480. Each element represents the temperature value of the corresponding spatial location, in °C, with an accuracy of 0.1 °C. The inertial measurement unit integrates a three-axis accelerometer and a three-axis gyroscope, outputting six-dimensional time-series signals of linear acceleration and angular velocity at a sampling frequency of 200 Hz. The global positioning system module provides three-dimensional position coordinates, velocity, and timestamp information, with a positioning accuracy better than 2.5 meters.

[0036] All sensor outputs are appended with a uniform timestamp, generated synchronously by a high-stability crystal oscillator with an error not exceeding ten microseconds. The raw sensor data is input to the subsequent processing unit in the form of a high-dimensional data stream. The data dimension depends on the sensor type: video streams are four-dimensional tensors (number of batches, time steps, height, width, number of channels), point clouds are three-dimensional tensors (number of batches, number of points, feature dimension), and multi-axis time-series signals are three-dimensional tensors (number of batches, time steps, number of axes).

[0037] In step S2, the asymmetric decoupling coding unit performs deep feature decoupling on the original sensor data. This unit is an end-to-end trained deep neural network model, and its structure includes three sub-units: a convolutional feature extraction sub-unit, a task-related representation generation sub-unit, and a privacy-sensitive representation generation sub-unit. The convolutional feature extraction sub-unit adopts a residual network architecture, consisting of 16 residual blocks connected in series. Each residual block contains two convolutional layers, two batch normalization layers, and a modified linear unit activation function.

[0038] For image input, the first convolutional kernel size is 7×7 with a stride of 2, resulting in 64 output channels. Subsequent residual blocks progressively expand the receptive field and increase the number of channels, ultimately outputting a shared feature map with dimensions of 8×8×512. For point cloud input, the convolutional feature extraction subunit is replaced with a point convolutional network, extracting local geometric features through dynamic graph construction and edge convolution operations. The task-related representation generation subunit and the privacy-sensitive representation generation subunit are connected after the shared feature map, both being three-layer fully connected networks.

[0039] The hidden layer neurons in the task-related representation generation subunit are 1024 and 512 respectively, while the output layer neurons are fixed at 256. No activation function is used at the ends; the output is a linear real-valued vector. The hidden layer neurons in the privacy-sensitive representation generation subunit are 256 and 128 respectively, while the output layer neurons are fixed at 64. Similarly, no activation function is used at the ends. The output vectors of the two subunits satisfy an orthogonality constraint in the latent space, meaning the absolute value of their inner product is less than a preset threshold of 0.01.

[0040] This orthogonality is achieved by introducing an orthogonal loss term during the training phase, defined as the square of the dot product of the task-related latent representation vector and the privacy-sensitive latent representation vector. The training of the asymmetric decoupled coding unit employs an adversarial mechanism, involving three discriminators: a task discriminator, a privacy discriminator, and a reconstruction discriminator.

[0041] The task discriminator is a classification network that receives task-related latent representations as input and outputs predictions for the corresponding AI task. Its loss function is cross-entropy loss. The privacy discriminator is a multilayer perceptron that attempts to reconstruct privacy attributes (such as facial recognition and geolocation) from the original sensor data from the task-related latent representations. Its loss function is mean squared error. The reconstruction discriminator receives a concatenated vector of task-related latent representations and privacy-sensitive latent representations, and attempts to reconstruct the original sensor data through a decoder network. Its loss function is also mean squared error. The total loss function is defined as: ; in, To reconstruct the loss, For mission losses, For the loss of the privacy discriminator, and The weighting coefficients are preset and range from 1.0 to 0.5. During training, the parameter update direction of the asymmetric decoupled coding unit is to minimize... and , while maximizing This forces privacy information to be completely isolated from privacy-sensitive potential representations.

[0042] After training, the weight parameters of the asymmetric decoupled coding unit are embedded in the hardware accelerator of the edge computing terminal. This hardware accelerator is a field-programmable gate array (FPGA) with a dedicated convolutional computation engine and matrix multiplication unit, supporting fixed-point operations to reduce power consumption. Raw sensor data is transferred from the sensor buffer to the on-chip memory of the hardware accelerator via a direct memory access channel, with processing latency controlled within 50 milliseconds.

[0043] In step S3, the task representation obfuscation transmission unit performs an irreversible transformation on the task-related latent representations. This unit includes a linear projection layer, a scalar quantization layer, and a data encapsulation module. The linear projection layer uses a fixed non-square matrix projection matrix. This matrix is ​​generated by a secure random number generator during system initialization. Its elements follow a standard normal distribution and are distributed as a system-level shared key to all edge terminals and cloud servers. (Task-related latent representation vector) After linear projection, the result is obtained Scalar quantization layer Each floating-point element in Perform 8-bit uniform quantization, the quantization formula is: ; in, and For vectors The global minimum and maximum values ​​are determined by historical statistics and updated periodically. (Quantized integer sequence) The data is fed into the data encapsulation module. This module generates a binary data frame. The frame header contains a 16-byte device unique identifier, an eight-byte nanosecond-level timestamp, and a four-byte cyclic redundancy check (CRC) code. The frame body is a 128-byte quantized integer sequence. The entire obfuscation task data packet is 156 bytes long, far smaller than the original video frame's 1.5 megabytes, achieving a compression ratio of over 10,000 times. After encapsulation, the data packet is sent to the cloud computing server via a low-power wide-area network (LPWAN) wireless communication module. Simultaneously, the privacy-sensitive potential representation vector is immediately overwritten with random noise after generation and released from memory, ensuring its lifetime does not exceed one millisecond, eliminating any possibility of storage or leakage.

[0044] In step S4, the task representation receiving unit of the cloud computing server parses and recovers the received obfuscated task data packets. First, the receiving unit verifies the cyclic redundancy check (CRC) code; if the verification fails, the data packet is discarded and a retransmission is requested. If the verification passes, the receiving unit extracts the device identifier and timestamp for log recording and data traceability, and then reads the 128-byte quantized integer sequence. The inverse quantization process converts each integer... Mapping back to floating-point values The formula is: ; Obtain the recovered projection vector Due to the projection matrix For a tall and thin matrix, its inverse mapping requires a generalized inverse matrix. Implementation. The task representation receiving unit pre-stores... Calculate the task-related latent representations of the recovered data. Although the recovered vector is different from the original... There are minor differences, but the task-specific inference unit is robust to this because the projection and quantization processes have been incorporated into end-to-end optimization during the training phase.

[0045] In step S5, the task-specialized inference unit receives... It performs artificial intelligence inference. This unit is a deep neural network trained in pairs with an asymmetric decoupled encoding unit, and its structure depends on the task type. In object detection tasks, the task-specialized inference unit employs an attention-based deconvolutional network. Input vector First, it is expanded into a four-dimensional tensor through a fully connected layer, and then gradually upsampled through four levels of deconvolution layers. After each level of deconvolution, a channel attention module is introduced to dynamically adjust the weights of each channel of the feature map.

[0046] The final output is a 3D tensor containing the target class probability distribution, confidence score, and bounding box coordinates. In the semantic segmentation task, the task-specialized inference unit adopts a U-shaped encoder-decoder structure. The encoding path consists of four downsampling blocks, and the decoding path consists of four upsampling blocks. Skip connections fuse the intermediate feature maps of the encoding path with the corresponding layers of the decoding path, ultimately outputting a pixel-level classification mask with the same resolution as the original image. After post-processing (such as non-maximum suppression), the inference result is returned to the requester through the application interface. The entire cloud inference process is executed on a GPU cluster, with a single inference cycle taking less than 100 milliseconds.

[0047] The system comprises edge computing terminals and cloud computing servers. Edge computing terminals are deployed on UAV onboard computers or IoT embedded devices, and their hardware platform includes a multi-core central processing unit (CPU), a field-programmable gate array (FPGA) hardware accelerator, various sensor interfaces, and a low-power wireless communication module. The CPU runs a real-time operating system, responsible for sensor data acquisition scheduling, hardware accelerator task triggering, and communication protocol stack management. The hardware accelerator stores the weight parameters of asymmetric decoupled coding units and exchanges data with the CPU through a high-speed serial interface. The cloud computing server is a distributed computing cluster, containing a task scheduler, a task representation receiving unit pool, and a task-specific inference unit pool.

[0048] The task scheduler routes the obfuscated task packets to the corresponding receiving unit instance based on the device identifier. After the receiving unit completes recovery, it will... A matching inference unit instance is sent in. The system maintains the connection between the edge terminal and the cloud through a heartbeat mechanism, supporting dynamic load balancing and failover.

[0049] This embodiment fully implements the entire process from data acquisition, privacy decoupling, obfuscated transmission to high-precision inference in the cloud, resolving the fundamental contradiction between privacy protection and model accuracy in edge computing. Through mandatory representation separation and unidirectional data flow design, it ensures that privacy data never leaves the edge device; through the powerful computing resources of the cloud, it guarantees the inference performance of complex artificial intelligence tasks; and through low-dimensional latent representation transmission, it significantly reduces communication overhead and terminal power consumption. This method is suitable for scenarios with stringent requirements for both privacy and accuracy, such as security monitoring, autonomous driving, and industrial inspection.

Claims

1. An artificial intelligence edge computing method for drones and IoT terminals, characterized in that, include: On the edge computing terminal side, raw sensor data is collected in real time through the data acquisition unit; The asymmetric decoupling coding unit built into the edge computing terminal processes the raw sensing data to generate mutually orthogonal task-related potential representations and privacy-sensitive potential representations. The task representation obfuscation transmission unit of the edge computing terminal performs irreversible obfuscation transformation and compression encoding on the task-related potential representation to generate an obfuscated task data packet to be transmitted, and sends the obfuscated task data packet to the cloud computing server through the wireless communication module; the privacy-sensitive potential representation is destroyed immediately after generation and is not stored or transmitted in any form. The task representation receiving unit of the cloud computing server receives and parses the obfuscated task data packet to recover the task-related potential representation. The task-specific inference unit of the cloud computing server receives the recovered task-related potential representations, performs a preset artificial intelligence inference task based on them, and outputs the final inference result.

2. The artificial intelligence edge computing method for drones and IoT terminals according to claim 1, characterized in that, The data acquisition unit includes at least one sensor configured on a drone or IoT terminal. The sensor is one or more of an image sensor, a lidar sensor, an infrared sensor, an inertial measurement unit, or a global positioning system module. The raw sensing data is a high-dimensional data stream containing timestamps, and its data format is video stream, point cloud data, thermal imaging matrix, or multi-axis time-series signal.

3. The artificial intelligence edge computing method for drones and IoT terminals according to claim 2, characterized in that, The asymmetric decoupling coding unit is a deep neural network model, whose structure includes a shared convolutional feature extraction subunit, and parallel task-related representation generation subunits and privacy-sensitive representation generation subunits connected to the back end of the convolutional feature extraction subunit. The convolutional feature extraction subunit adopts a residual network structure to perform multi-layer convolution and pooling operations on the input raw sensor data to extract a high-dimensional shared feature map. The task-related representation generation subunit and the privacy-sensitive representation generation subunit are both composed of multi-layer fully connected networks, which respectively map the shared feature map into task-related latent representation vectors and privacy-sensitive latent representation vectors of predetermined dimensions.

4. The artificial intelligence edge computing method for drones and IoT terminals according to claim 3, characterized in that, The training process of the asymmetric decoupling coding unit adopts an adversarial training mechanism. This training process involves a generator, namely the asymmetric decoupling coding unit, and three independent discriminator networks: a task discriminator, a privacy discriminator, and a reconstruction discriminator. The overall training objective of the asymmetric decoupled coding unit is defined by a composite loss function, the mathematical expression of which is: Total Loss = Reconstruction Loss + First Weighting Coefficient Multiplied by Task Loss - Second Weighting Coefficient Multiplied by Privacy Adversarial Loss; wherein, the reconstruction loss is used to measure the difference between the data reconstructed using the task-related latent representation and the privacy-sensitive latent representation and the original sensing data, and its calculation method is mean squared error; the task loss is generated by the task discriminator and is used to measure the accuracy of completing a specific artificial intelligence task using the task-related latent representation, and its calculation method is cross-entropy loss; The privacy adversarial loss is generated by the privacy discriminator and is used to measure the success rate of recovering privacy information from the task-related latent representations. The asymmetric decoupling coding unit is committed to minimizing the reconstruction loss and task loss during training, while maximizing the privacy adversarial loss.

5. The artificial intelligence edge computing method for drones and IoT terminals according to claim 4, characterized in that, The task representation obfuscation transmission unit includes a linear projection layer, a scalar quantization layer, and a data encapsulation module. The linear projection layer contains a fixed, non-square projection matrix that linearly maps the task-related potential representations from a first-dimensional space to a lower-dimensional second-dimensional space. The scalar quantization layer converts each floating-point value in the vector after linear projection into a fixed-point integer value according to a preset quantization bit width. The data encapsulation module combines the quantized integer sequence with a frame header containing a device identifier, timestamp, and data checksum into a binary-format obfuscated task data packet.

6. The artificial intelligence edge computing method for drones and IoT terminals according to claim 5, characterized in that, The projection matrix used by the linear projection layer is a random Gaussian matrix. This matrix is ​​generated and fixed during the system initialization phase and distributed to all edge computing terminals and cloud computing servers as a shared key within the system. The quantization bit width of the scalar quantization layer is set to eight bits, mapping each floating-point value to an integer range of 0-255.

7. The artificial intelligence edge computing method for drones and IoT terminals according to claim 6, characterized in that, After receiving the obfuscated task data packet, the task representation receiving unit of the cloud computing server first performs a data check code verification to confirm the integrity of the data packet. Subsequently, it extracts the quantized integer sequence and restores it to a floating-point vector through an inverse quantization process. Finally, it applies the generalized inverse matrix of the projection matrix to reverse map the floating-point vector from the second-dimensional space back to the first-dimensional space, thereby restoring a vector that is numerically similar to the task-related potential representation generated by the edge computing terminal, and transmits it to the task-specialized inference unit.

8. The artificial intelligence edge computing method for drones and IoT terminals according to claim 7, characterized in that, The task-specific inference unit is a deep neural network decoder trained in pairs with the asymmetric decoupled encoding unit. In the object detection task, the task-specific inference unit adopts a multi-layer deconvolutional network structure based on an attention mechanism to progressively upsample and decode the input low-dimensional latent representation into an output tensor containing the target category, confidence score, and bounding box coordinates. In the semantic segmentation task, the task-specific inference unit adopts an encoder-decoder structure to restore the latent representation into a pixel-level segmentation mask with the same resolution as the original sensor data through a series of upsampling and skip connection operations.

9. The artificial intelligence edge computing method for drones and IoT terminals according to claim 8, characterized in that, The dimension of the task-related latent representation is set to 256, and the dimension of the privacy-sensitive latent representation is set to 64. The terminal layers of the task-related representation generation subunit and the privacy-sensitive representation generation subunit do not use any activation function to output a linear latent space distribution.

10. The artificial intelligence edge computing method for drones and IoT terminals according to claim 9, characterized in that, The asymmetric decoupling coding unit is implemented by a hardware accelerator, which is a field-programmable gate array or an application-specific integrated circuit; the neural network weight parameters of the asymmetric decoupling coding unit are stored in the read-only memory of the hardware accelerator. The processor of the edge computing terminal feeds raw sensor data to the hardware accelerator through a direct memory access channel.

Images