A multi-modal living body anti-counterfeiting and end-side AI anti-fraud identification system based on TEE

Abstract

The application discloses a multi-modal living body anti-fake and end-side AI anti-fraud identification system based on TEE, and belongs to the technical field of terminal AI security and biological recognition anti-fraud. In view of the technical difficulties that the existing single-mode living body recognition is easy to be counterfeited, the cloud anti-fraud privacy is leaked, the AI deep counterfeiting cannot be recognized, the biological data storage is not safe, and the scene verification is not hierarchical, a multi-dimensional multi-modal living body fusion verification system is constructed, and various types of counterfeiting attacks are eliminated by combining 3D face, subcutaneous vein, eyeball micro-motion and micro-expression physiological characteristics; local offline AI synthesized voice, false short message and phishing link identification are realized through an end-side lightweight anti-fraud large model; biological feature encryption isolation storage is realized by relying on a TEE hardware security partition, and the safety of user privacy is ensured; meanwhile, a scene adaptive hierarchical verification mechanism is designed, and differentiated security strategies are implemented for ordinary unlocking, small amount operation and large amount financial transaction, and multiple verification and anti-fraud re-check are forced in high-risk scenes. The application does not need cloud data uploading, has high privacy security, can comprehensively defend new AI fraud means, constructs a terminal payment level security barrier, and can be widely applied to smart phones, financial terminals, government and enterprise security equipment, and has high technical barrier and industrial value.

Description

Technical Field

[0001] This invention relates to the fields of mobile terminal biometric security, edge AI anti-fraud, and trusted execution environment (TEE) hardware encryption technology. Specifically, it relates to multimodal liveness fusion anti-counterfeiting algorithms, offline AI deepfake recognition, TEE isolated biometric data storage, and scenario-adaptive hierarchical security verification technology, which are applicable to identity verification and anti-fraud protection scenarios for smartphones, financial terminals, government and enterprise security terminals, and operator-customized terminals. Background Technology

[0002] Current mobile terminal identity authentication and telecommunications anti-fraud systems suffer from several key technical deficiencies. First, traditional terminals generally employ a single 2D face or fingerprint monomodal recognition method, which is weakly protected and easily bypassed by high-definition photos, copied videos, 3D-printed masks, silicone prostheses, and AI-generated face-swapped videos. They also lack deep physiological liveness verification capabilities. Second, existing anti-fraud systems largely rely on cloud-based keyword matching mechanisms, requiring the uploading of user call, SMS, and image data to cloud servers. This not only poses a privacy risk but also fails to identify new deepfake fraud methods such as AI-synthesized speech, AI digital human face-swapping, fake customer service SMS messages, and fake phishing links, resulting in high recognition latency and serious false positives and false negatives.

[0003] Meanwhile, existing biometric data is mostly stored in the terminal's rich operating system partition, which is weakly encrypted. System vulnerabilities and malicious applications can illegally read and steal users' biometric templates, and once biometric data is leaked, it will cause permanent identity security risks. In addition, the existing terminal security verification mechanism uses a uniform verification standard, which cannot distinguish between different risk scenarios such as ordinary unlocking, small-amount operations, large-amount transfers, and privacy access. The protection level for high-risk financial operations is insufficient, which can easily lead to security problems such as fund theft and telecommunications fraud. Existing technologies cannot simultaneously achieve high-strength liveness detection, offline anti-fraud on the terminal side, hardware-level privacy protection, and scenario-based security control, and have significant technical shortcomings. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies, such as easy cracking of single-modal anti-counterfeiting, high privacy risks of cloud-based anti-fraud, lack of deepfake identification capabilities, insecure storage of biometric data, and lack of hierarchical control of security verification. This invention provides a multimodal liveness detection and edge AI anti-fraud identification system based on TEE, which realizes payment-level security protection for terminal devices and AI anti-fraud capabilities in all scenarios.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The system of this invention is divided into three layers: a rich operating system application layer, a lightweight AI inference layer on the edge side, and a TEE trusted hardware security layer. The three layers are unidirectionally isolated from each other. All highly sensitive operations, biometric comparisons, and security verification logic are executed within the TEE security zone, eliminating the risk of data leakage and tampering.

[0006] This invention comprises four core technology modules: a multimodal liveness fusion anti-counterfeiting module, an edge-side offline AI anti-fraud identification module, a TEE hardware isolation encrypted storage module, and a scene-adaptive hierarchical security verification module.

[0007] Furthermore, the multimodal liveness fusion anti-spoofing module simultaneously collects 3D facial depth point cloud data, subcutaneous vein texture features, eye micro-movement temporal trajectory, and facial micro-expression dynamic features of the terminal user to construct a multi-dimensional physiological liveness feature database. Through a multimodal feature fusion neural network, each dimension feature is independently scored and weighted to output a comprehensive liveness confidence score. Differentiated discrimination thresholds are set for four types of attack scenarios: static photos, dynamic videos, physical masks, and AI face swapping. If a single feature does not meet the liveness condition, verification is rejected, completely eliminating various identity fraud attacks.

[0008] Furthermore, the on-device offline AI anti-fraud identification module deploys a self-developed lightweight anti-fraud model on the local inference unit of the terminal, without relying on cloud servers and without uploading user privacy data throughout the process; it analyzes call voice, SMS text, and page link information locally in real time, and identifies AI-synthesized fraudulent voice through timbre spectrum, breathing rhythm, and human voice dynamic features, and identifies fake customer service, public security, procuratorate, and bank information and phishing fake links through semantic features, script templates, and link features, thus achieving offline real-time anti-fraud risk assessment.

[0009] Furthermore, the TEE hardware-isolated encrypted storage module encrypts all collected biometric data using the TEE's internal national cryptographic algorithm, generating irreversible biometric templates that are stored only in the TEE's independent secure partition. Operating system kernels, third-party applications, and malicious programs have no read, write, or decryption permissions. Biometric comparison and verification operations are completed entirely within the TEE's trusted computing unit, outputting only the final verification result. The original biometric data never leaves the secure area, achieving hardware-level privacy protection.

[0010] Furthermore, the scenario-adaptive hierarchical security verification module has a built-in terminal operation scenario risk classification rule library, which automatically identifies the current operation behavior and matches the corresponding verification strategy: Level 1 low-risk scenarios use single-factor liveness verification, Level 2 medium-risk scenarios use dual-modal joint verification, and Level 3 high-risk scenarios such as large-amount transfers, card binding, and privacy access are subject to mandatory quadratic full-scale liveness verification + AI anti-fraud risk secondary review; when the anti-fraud module identifies fraud risk, it automatically upgrades to the highest security verification level and triggers a risk warning. Beneficial effects

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. Comprehensive defense against new AI fraud and liveness detection attacks. This invention integrates multi-dimensional liveness features such as 3D face, veins, micro-eye movements, and micro-expressions, which can completely resist fake attacks such as photos, videos, masks, 3D prosthetics, and AI face swapping; the offline AI model on the device accurately identifies AI-synthesized speech, fake text messages, and phishing links, covering all current mainstream new types of telecommunications fraud.

[0012] 2. Offline processing on the device, zero privacy leakage. All liveness detection, anti-fraud identification, and data comparison are completed locally on the device. User biometric information, call data, and SMS data do not need to be uploaded to the cloud, fully complying with the requirements of the Personal Information Protection Law and the Data Security Law, eliminating the risk of privacy leakage at the source.

[0013] 3. TEE hardware-level security barrier with extremely high protection level. Relying on the trusted execution environment of hardware to achieve physical isolation storage of biometric data, it eliminates system vulnerabilities and malicious APP theft and tampering of biometric templates, builds a bank-level payment security foundation, and solves the security shortcomings of traditional terminal biometric authentication.

[0014] 4. Tiered dynamic verification, balancing security and usability. Simplified verification in low-risk scenarios ensures a good user experience, while mandatory multi-factor verification is enforced in high-risk financial scenarios. Simultaneously, AI-powered anti-fraud measures are dynamically adjusted to optimize security strategies, achieving a balance between security and user experience.

[0015] 5. High versatility and wide range of application scenarios. This invention can be mass-produced and adapted to smartphones, financial POS terminals, government and enterprise confidential terminals, and operator-customized security equipment, possessing extremely high industrial commercial value and technological barriers. Detailed Implementation

[0016] The present invention will be further described in detail below with reference to specific embodiments.

[0017] This invention discloses a multimodal liveness detection and edge AI anti-fraud identification system based on TEE, the workflow of which is as follows: Step 1: The terminal starts the identity verification or risk monitoring task. The multimodal acquisition module simultaneously acquires the user's 3D face depth data, subcutaneous vein texture, eye micro-movement trajectory, and facial micro-expression temporal features to complete the acquisition of multi-dimensional live original features.

[0018] Step 2: The multimodal fusion model scores the four types of features for liveness detection and calculates the weighted comprehensive confidence score. If the score is lower than the preset liveness threshold, the operation is directly blocked and a spoofing attack risk warning is output.

[0019] Step 3: The lightweight AI anti-fraud model on the device monitors the device's call, SMS, and network access behavior in real time, analyzes audio, video, and text content offline, identifies AI-synthesized speech, fake fraudulent scripts, and forged links, and generates risk labels.

[0020] Step 4: All legitimate biometric data is encrypted using the TEE's internal national cryptographic algorithm and stored in a dedicated secure partition of the TEE, which cannot be accessed or decrypted by the system layer or application layer.

[0021] Step 5: The system automatically matches a tiered verification strategy based on the current operation scenario: ordinary unlocking uses single-factor verification, small-amount operations use dual-modal verification, and high-risk scenarios such as large-amount transfers and privacy access are subject to mandatory full multimodal verification plus anti-fraud risk review.

[0022] Step 6: If the risk of AI-generated fraud is detected, the system will automatically upgrade the security verification level, lock sensitive operations and issue a pop-up warning to complete the anti-fraud protection.

[0023] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, optimizations, or improvements made within the spirit and principles of the present invention shall fall within the scope of protection of the present invention. Attached Figure Description

[0024] Figure 1 This is a three-layer overall architecture diagram. It shows the system's top-down three-layer unidirectional isolation architecture, consisting of the rich operating system application layer, the edge-side lightweight AI inference layer, and the TEE trusted hardware security layer, as well as the deployment relationship of the four core modules. Figure 2 This is a complete workflow diagram of the multimodal liveness fusion anti-counterfeiting module of the present invention. It is used to demonstrate the complete liveness verification logic, including simultaneous acquisition of multiple physiological features, independent scoring of single-channel features, multi-threshold attack discrimination, and fusion confidence output. Figure 3 is a logic diagram of the local offline risk discrimination of the offline AI anti-fraud identification module on the edge side of the present invention. It shows the whole process of local feature extraction from multiple sources of terminal voice, SMS and link data, offline inference of lightweight model and output of fraud risk label, reflecting the offline characteristics of not uploading original privacy data. Figure 4 is a timing block diagram of the linkage between the TEE hardware isolation encrypted storage and the scenario adaptive hierarchical security verification of the present invention. It shows the biometric hardware encryption isolated storage process, the three-level risk scenario adaptive verification strategy, and the linkage mechanism for triggering security level upgrades due to anti-fraud risks.

Claims

1. A multimodal liveness detection and edge AI anti-fraud identification system based on TEE, characterized in that, include: Multimodal liveness fusion anti-counterfeiting module, edge-side offline AI anti-fraud recognition module, TEE hardware isolation encrypted storage module, and scene-adaptive hierarchical security verification module; The multimodal liveness fusion anti-counterfeiting module is used to collect multi-dimensional liveness physiological features such as 3D face depth point cloud, subcutaneous vein texture, eye micro-movement trajectory, and facial micro-expression. It uses a multimodal fusion neural network to determine the liveness confidence and intercept static, dynamic, and AI forgery attacks. The on-device offline AI anti-fraud identification module is used to deploy a lightweight anti-fraud model locally on the terminal, and to analyze call voice, SMS text and network connection information offline to identify AI-synthesized fraudulent voice, impersonation scripts and fake phishing information. The TEE hardware isolation encrypted storage module is used to encrypt biometric data using the TEE's internal national cryptographic algorithm to generate an irreversible template, which is then stored in an independent secure partition of the TEE, achieving isolation and inaccessibility between the system layer and the application layer. The scenario-adaptive hierarchical security verification module is used to adaptively match single-factor, dual-factor, and multi-factor full-scale liveness verification strategies according to the risk level of terminal operation, and link the anti-fraud model for secondary risk review in high-risk scenarios.

2. The system according to claim 1, characterized in that, The multimodal liveness fusion anti-counterfeiting module sets independent discrimination thresholds for static photo attacks, dynamic video re-enactment attacks, physical mask spoofing attacks, and AI face-swapping synthesis attacks. If any single-dimensional feature fails the verification, it is directly judged as an illegal forgery of identity.

3. The system according to claim 1, characterized in that, The on-device offline AI anti-fraud identification module completes data parsing and risk assessment entirely on the local terminal, without uploading the user's original voice, text, or biometric privacy data, and only outputs risk assessment result labels.

4. The system according to claim 1, characterized in that, In the TEE hardware-isolated encrypted storage module, biometric comparison and encryption operations are all performed within the TEE trusted computing unit. The original biometric data does not flow out of the secure partition, and the rich operating system and third-party applications have no read, write, or decryption permissions.

5. The system according to claim 1, characterized in that, The scenario-adaptive hierarchical security verification module divides the risk strategy into three levels: Level 1 low-risk scenarios use single-modal verification, Level 2 medium-risk scenarios use dual-modal joint verification, and Level 3 high-risk financial operations are subject to mandatory quadratic full-scale liveness verification; when fraud risk is detected, the highest protection level is automatically upgraded and an early warning interception is triggered.

Images