Data-Secure Artificial Intelligence Service with Hierarchical AI on the Edge

Abstract

The invention relates to an edge computing-based artificial intelligence service that protects user data privacy, minimizes latency, and can operate independently of an internet connection.

Description

1 TARIFF Data-Secure Artificial Intelligence Service with Hierarchical AI on the Edge Technical Area 5 The invention protects user data privacy, minimizes latency, and improves internet connectivity. an edge computing-based system that can operate independently of its connection It is related to artificial intelligence services. State of the Art Today, smart home devices handle a large portion of services such as parental controls and network management. This is offered through cloud-based assistants. This approach addresses data privacy concerns, It has disadvantages such as latency issues and dependence on the cloud. The current cloud is 15. The biggest shortcoming of the assistants lies in the way the data is processed. In these systems... User voice recordings, messages, and contextual information are generally in plain text. The data is processed and stored on the cloud provider's servers. In this case, the cloud... There is a possibility that the provider, data center operators, or unauthorized persons may access this data. Time is a factor. Whether user data is used for model training is 20. or transparency regarding how long it is stored is limited. Advanced security measures However, centralized servers are targets for large-scale data breaches. The costs of cloud assistants become rapidly uncontrollable as usage increases. This is possible. Furthermore, cloud assistants have an absolute dependence on internet connectivity. 25 Due to its nature, it has a fragile structure. When the internet connection is interrupted or the cloud... When there is an interruption in the service, the assistant becomes completely inoperable. This situation, This is a major disadvantage for critical functions such as home automation or security. On the other hand, the fact that each command has to go back and forth to a data center, especially in voice interactions, can take 30 minutes. This causes a noticeable delay. This negatively impacts the user experience. It has an impact. In particular, the initial response in systems that have not been used for a long time and are being brought back online. The duration may increase. Local and global data regulations (KVKK, GDPR) and geopolitical tensions make it difficult to transfer data within the country. 35 This makes it increasingly difficult to process data on cloud servers outside of Turkey. Strict regulations like GDPR... 2 For companies or organizations subject to regulations, user data is handled by the cloud provider. Transferring the servers to the country where they are located may pose a legal obstacle. Platform owners like Google and Apple have third parties who are dependent on their ecosystems. They have the authority to restrict or completely cut off access to assistants. This affects business continuity. This poses a serious risk. Data is being processed outside of your control. 5 This makes auditing and identifying security vulnerabilities more difficult. In conclusion, while current cloud-based assistants offer ease of use, ISPs' service requirements remain the same. It has shortcomings that contradict its core values ​​such as quality, data privacy, and cost control. At this point, the local AI assistant (Edge AI) model addresses these shortcomings by providing full control over the data. It offers a strategic infrastructure that eliminates [problems]. Due to the negative aspects described above and the current solutions regarding the issue... Due to its shortcomings, it has become necessary to make improvements in the relevant technical field. Figures that will help understand the invention. Figure 1 shows the general architecture of the system that is the subject of the invention. Explanation of Part References 100. User equipment 20 200. Customer facility equipment 300. Edge calculation node 400. Cloud server Detailed Description of the Invention 25 In this detailed explanation, the preferred configurations of the invention are not merely for better understanding the subject. It is intended to facilitate understanding and will not impose any limiting effects. The invention, for the user protecting the privacy of your data, minimizing latency, and from your internet connection an edge computing-based artificial intelligence that can operate independently 30 It is related to the service. A schematic view of the system described in the invention is provided. From the schematic view, it can be seen that the telephone, User equipment such as computers, tablets (100) as well as smart home devices, AI modules, Running on the ISP's next-generation modem / gateway, the device has a lightweight NLP model 35 Customer facility equipment with natural language processing and rule-based automation engine. 3 (200), building base stations, routers used by customers to connect to the internet, access points consisting of switches and servers, on an apartment or street basis The small edge servers deployed (MEC – Multi-access Edge Computing) are larger multiple models (e.g., anomaly detection, object recognition) can be run in the same building The device can provide services to the subscriber, including all voice commands, in-home traffic analysis, and camera images. or some models that can remain at the edge of the neighborhood, upon user request (e.g., a new language) Edge computing node (300) that can be updated via secure channel (support), user data By anonymizing the information, only model updates, usage statistics (anonymous), and The configuration is managed via the cloud, and personal data absolutely does not go to the cloud. The server contains (400). 10 The operating principle of the system described in the invention can be summarized as follows. The user gives a voice command or enters a request via the mobile application (100). AI-powered modem / gateway (200), a continuously running low-power wake word engine 15 It monitors the sound. When the word "wake up" is detected, the voice recording does not exit the device; it is a local one. Noise reduction and speech recognition pass through the preprocessing layer. At this stage, simple A lightweight language model (SLM) for commands (e.g., “open child profile”) is directly provided to the customer facility. It can be operated on equipment (200). If the command is more complex (e.g., “check all cameras in the house, for movement in the last hour) is there any?”) or if the customer's facility equipment (200) exceeds the processing capacity, request Anonymized edge calculation node (300) at the neighborhood or building level It is redirected. This server handles larger models (image processing, long-context natural language). (understanding). At this stage, the user's personal data (voice recording, image) is included. It is temporarily processed on the computing node (300) and immediately deleted. Never It is not transmitted to the ISP's central cloud or to third parties. AI model in edge computing node (300) with command local context information It enriches it. For example, when told to "turn off the living room light," the system first activates home automation 30 It controls the device topology (Matter, Zigbee) on the network. It is proactive, such as ensuring child safety. In these scenarios, anomaly models constantly run in the background to analyze traffic behavior. It automatically applies the rules. As a result of the decision, the system network policy change (QoS, content filtering, device-based speed limiting) 35 The data is transmitted directly to the software-defined network (SDN) engine of the customer facility equipment (200). IoT 4 Device commands are sent to devices via the local hub (e.g., Zigbee coordinator); over the internet It works even without a connection. User notification voice response, mobile app (100) push or This is done via the dashboard. Thousands of edge computing nodes / customer facility equipment servers (300) only model 5 Weight updates are anonymously and encrypted to a central cloud (400) The raw data is never sent to the coordinator. This allows a generalized global model to be learned while the raw data is never used. It is not aggregated. Federationd learning is used for model updates. The operating principle of the system described in the invention can be summarized as follows. The user gives a voice command or enters a request via the mobile application (100). AI-powered modem / gateway (200), a continuously running low-power wake word engine It monitors the sound. When the word "wake up" is detected, the voice recording does not exit the device; it is a local one. Noise reduction and speech recognition pass through the preprocessing layer. At this stage, simple A lightweight language model (SLM) for commands (e.g., “open child profile”) is directly provided to the customer facility. 15 It can be operated on equipment (200). If the command is more complex (e.g., “check all cameras in the house, for movement in the last hour) is there any?”) or if the customer's facility equipment (200) exceeds the processing capacity, request Anonymized edge calculation node (300) 20 at neighborhood or building level It is redirected. This server handles larger models (image processing, long-context natural language). (understanding). At this stage, the user's personal data (voice recording, image) is placed aside. It is temporarily processed on the computing node (300) and immediately deleted. Never It is not transmitted to the ISP's central cloud or to third parties. AI model in edge computing node (300) with command local context information It enhances the experience. For example, when told to "turn off the living room lights," the system first activates home automation. It controls the device topology (Matter, Zigbee) on the network. It is proactive, such as ensuring child safety. In these scenarios, anomaly models constantly run in the background to analyze traffic behavior. It automatically applies the rules. 30 The decision resulted in changes to the system network policy (QoS, content filtering, device-based speed limiting). The data is transmitted directly to the software-defined network (SDN) engine of the customer facility equipment (200). IoT Device commands are sent to devices via the local hub (e.g., Zigbee coordinator); over the internet It works even without a connection. User notification voice response, mobile app (100) push or 35 This is done via the dashboard. Thousands of edge computing nodes / customer facility equipment servers (300) only model Weight updates are anonymously and encrypted to a central cloud (400) The raw data is never sent to the coordinator. This allows a generalized global model to be learned while the raw data is never used. It is not aggregated. Federationd learning is used for model updates.

Claims

6 REQUESTS 1. Protecting user data privacy, minimizing latency, and the internet Edge computing, which can operate independently of its connection. It is an AI-based service with the following features: 5  smart home in addition to user equipment such as phones, computers, tablets (100) devices,  The AI ​​module runs on the next-generation modem / gateway provided by the ISP, The device features a lightweight NLP model (natural language processing) and a rule-based automation engine. Customer facility equipment found (200), 10  Building base stations are used by customers to connect to the internet. apartment buildings with access points consisting of routers, switches, and servers or small edge servers deployed on a street-by-street basis (MEC – Multi-access) Edge Computing) for larger models (e.g., anomaly detection, object recognition) capable of operating, serving multiple subscribers in the same building, all voice 15 Commands, home traffic analysis, camera footage on the device or at the neighborhood edge. Some models remain secure upon user request (e.g., support for a new language). Edge calculation node (300) that can be updated from the channel,  User data is anonymized and used only for model updates and usage. Statistics (anonymous) and configuration managed via the cloud, personal data 20 cloud server that definitely does not go to the cloud (400) It includes.

2. An artificial intelligence service that complies with Claim 1, and whose feature is;  The user gives a voice command or enters a request via the mobile application (100). AI 25 Supported modem / gateway (200), continuously running low power wake word The (wake word) engine monitors the sound. When the wake word is detected, the sound recording is activated. It doesn't leave the device; it uses local noise reduction and speech recognition preprocessing. It passes through the layer. At this stage, it's lightweight for simple commands (e.g., "open child profile"). a language model (SLM) directly on customer facility equipment (200) 30 can be run.  If the command is more complex (e.g., “check all cameras in the house, one last one Is there any movement per hour?”) or customer facility equipment (200) operation If it exceeds capacity, demand is anonymized and reported on a neighborhood or building basis. Edge computing node (300) is directed. This server handles larger models 35 (It includes image processing and long-context natural language comprehension. At this stage...) 7 User's personal data (voice recording, image) edge computing node (300) It is processed temporarily on their end and immediately deleted. It never reaches the ISP's central office. It is not transferred to the cloud or to third parties.  AI model in edge computing node (300) with local context information of the command It enriches it. For example, when told to "turn off the living room light," the system first selects home 5 It controls the device topology (Matter, Zigbee) in the automation network. Child In proactive scenarios such as security, anomalies are constantly running in the background. The models automatically enforce rules by analyzing traffic behavior.  As a result of the decision, system network policy changes (QoS, content filter, device-based speed limiting) direct customer facility equipment (200) software defined network 10 (SDN) commands are transmitted to the engine. IoT device commands are sent via a local hub (e.g., Zigbee) (coordinator) is sent to the devices; it works even without an internet connection. User notification via voice response, mobile app (100) push or dashboard It is done.  Thousands of edge computing nodes / customer facility equipment from the server (300) 15 Only model weight updates are sent anonymously and encrypted to the cloud. (400) is sent to a central coordinator. Thus, a generalized global model Raw data is never collected during the learning process. A federation is used for model updates. Federationd learning is used.  It includes the steps of the process. 20

Images