AN ARTIFICIAL INTELLIGENCE LEARNING SYSTEM AND METHOD WITH AN ONLINE CORRECTION MECHANISM TO PROTECT THE INTELLECTUAL PROPERTY OF THE MODEL.

Abstract

This invention relates to the technical field of artificial intelligence security and the protection of model intellectual property (IP). The invention specifically aims to provide protection against threats such as model stealing and reverse engineering, particularly in machine learning and deep learning-based systems.

Description

1 TARIFF ONLINE CORRECTION TO PROTECT MODEL INTELLECTUAL PROPERTY. AN ARTIFICIAL INTELLIGENCE LEARNING SYSTEM AND METHOD WITH ITS OWN MECHANISM Technical Field to Which the Invention Relates This invention falls under the technical field of AI safety and the protection of model intellectual property (IP) 5. It is related to. The invention, particularly in machine learning and deep learning-based systems, Protection against threats such as model stealing and reverse engineering. It aims to provide this. State of the Art In the field of artificial intelligence, the knowledge accumulated from large and complex models (teacher models) is 10 a method of distilling information to transfer it to smaller and more efficient models (student models) It is widely used. This allows for the reduction of computational costs while the model... performance can be maintained. However, this process depends on the teacher model's outcomes. (Their predictions) are obtained by attackers, thus compromising the internal structure of the model or This brings with it the risk of their behavior being copied, in other words, the threat of model theft. 15 It brings. In order to solve this problem of intellectual property leakage, various approaches to the prior art have been developed. Defense mechanisms have been developed. One of them is the 2021 report by Ma and colleagues. It is the "Undistillable Teacher" approach that he proposed in 2008. In this method, the teacher model, The student model uses 20 probability distributions to make it difficult for the model to obtain accurate information. It deliberately distorts things. However, the fundamental weakness of this approach is that it makes information transfer more difficult. it also reduces the overall accuracy of the model and the learning process The risk is that it could completely disrupt everything. Kundu and colleagues' 2021 study, titled "Skeptical Student," 25 to help obtain limited information from distorted teacher outputs It uses classifiers and a self-distillation strategy. However, this The method adequately captures the hidden and valuable information in the teacher model's predictions. It fails to capture it and loses a significant portion of the information. Common shortcomings and weaknesses of these applications in the known state of the art. These are: 30 2 Misleading Information Transmission: Perturbed or misleading teacher predictions, any It is transmitted directly to the student without any correction mechanism, which means the student This leads to learning faulty patterns. Lack of Feedback: Information flow is one-way, from teacher to student. The student's a learning situation or performance, a factor affecting the teacher or the information transfer process. It does not generate feedback. This situation prevents the creation of an adaptive learning environment. obstacles. Static Correction: In current methods, real data is used to correct corrupted information. There is no time-bound and dynamic mechanism. Therefore, the learning process... Throughout this process, the transmission of incorrect information continues. 10 Inadequate Intellectual Property Protection: These efforts to hinder the distillation of knowledge, It cannot completely prevent information leakage even from perturbed outputs, and the IP model... It is unable to provide complete protection. The fundamental technical problem that arises as a result of these shortcomings is "perturbed". the inability to efficiently extract reliable information from teacher predictions" and this 15 As a result, both the student model's performance and the model's intellectual development decreased. The inability to fully protect property rights. Brief Description and Objectives of the Invention The primary goal of this invention is to protect the intellectual property rights of artificial intelligence models while simultaneously a system and method that enables high-performance data transfer in time 20 to present. Another purpose of the invention is the "undistillable teacher". distorted or misleading predictions produced by models in real time The goal is to provide a structure that can correct this. Another aim of the invention is to prevent the student model from being merely a passive recipient of information. 25 preventing, actively participating in the learning process and the information transfer mechanism The goal is to create a two-way learning cycle that generates feedback for improvement. 3 Another aim of the invention is to uncover the "dark" elements hidden in perturbed teacher outputs. The student model learns by recovering "knowledge" without any loss. The goal is to increase its efficiency. Another objective of the invention is model safety (IP protection) and model performance. By striking a balance between (truth) and innovation, these five innovative solutions simultaneously meet both needs. It aims to create an online knowledge distillation framework. Explanation of the Figures Figure 1: A block diagram showing the general architecture of the system described in the invention. Figure 2: A flowchart showing the working steps of the method described in the invention. Explanation of References in Figures 10 1. Input Data Unit 2. Teacher Model 3. Corrective Module 4. Student Model 5. Output Unit 15 6. Input Normalization Layer 7. Feature Correction Layers 8. Corrected Forecast Output 9. Input Fusion Layer 10. Prediction Layer 20 11. Feedback Generator 1001. Input Data Unit (1) by raw input data to teacher model (2) and student model (4) simultaneous transmission 1002. Teacher Model (2) intentionally for the purpose of intellectual property protection Generating perturbed estimates 25 1003. Input Normalization Layer (6) located within Corrective Module (3) normalization of distorted estimates from the teacher model (2) made ready for processing 1004. At least one Feature Correction Layer (7) located within Correction Module (3) by, perturbed information through nonlinear transformations 30 correction and retrieval of meaningful information 4 1005. Corrected and reliable probability distributions by the Corrective Module (3). Generating the Corrected Prediction Output (8) and transferring it to the Student Model (4) 1006. By the Input Fusion Layer (9) located within the Student Model (4), Input Raw data from Data Unit (1) and corrected data from Corrective Module (3) Combining the predictions 5 By the Prediction Layer (10) within the Student Model (4), 1007. Generating final class predictions based on fused data. 1008. Feedback Generator (11) analyzes the outputs produced by the Student Model (4). by generating a feedback signal and sending this signal to the Corrective Module (3) transmission 10 1009. Output Unit (5) by Student Model (4) and / or Corrective Module (3) presenting the final results from the source to the user or external systems. Detailed Description of the Invention This invention is known as the "Corrector Student Framework". called, online 15 developed to protect the intellectual property of artificial intelligence models. It is a system and method with a correction mechanism. The invention is intentionally damaged. Information from a Teacher Model (2) that generates predictions, a Corrective Module (3) corrected through and transferred to a Student Model (4) and this process is a feedback It is based on continuous optimization through a cycle. The process involves an Input Data Unit (1) inputting raw data (e.g., images, texts) into the system. It starts with feeding. This data will be protected as intellectual property for the Teacher Model (2) This information is also transmitted to the Student Model (4) who will learn it. The Teacher Model (2) is similar to the "undistillable" structure in the known state of the technique. In this way, deliberately to prevent the model's behavior from being copied from the outside. It produces estimates that contain perturbed or distorted probability distributions. This 25 Incorrect predictions, when used directly, will negatively impact the learning process. The invention is directed to the Corrective Module (3), which is the main innovation of the system. Corrective Module (3) processes the distorted predictions from the Teacher Model (2) It is the central component that transforms the data into reliable information. The first layer within this module is a The Input Normalization Layer (6) is used to normalize the incoming noisy prediction distributions to a certain 30 It prepares the data for processing by bringing it to scale. Then, the normalized data, at least one The feature is passed to the Feature Correction Layer (7). These layers are preferably convolutional neural networks. (CNN) based architecture and ReLU for capturing nonlinear relationships. It has activation functions. Thanks to this structure, perturbations are eliminated. is removed and hidden meaningful information ("dark knowledge") in the predictions is recovered. Corrected. The clean and reliable probability distributions obtained as a result of the correction process, It is produced as an Adjusted Prediction Output (8) and transferred to the Student Model (4). 5 The Student Model (4) learns these corrected estimates from the Corrective Module (3). It uses it in the process. Another innovative feature of the invention is that it is within the Student Model (4). This is the Input Fusion Layer (9). This layer receives data from the Input Data Unit (1). original raw data and corrected teacher information from the Corrective Module (3) It combines. This feature fusion allows the student to understand both real data patterns and 10 learning by enabling the teacher to simultaneously learn from their refined knowledge base It increases its efficiency. The Student Model (4), which processes this fused information, is a Prediction It generates its own final class predictions through layer (10). One of the invention's most distinctive features is its online feedback system, which breaks the one-way flow of information. It is a feedback mechanism. The outputs produced by the Student Model (4) are integrated into the system as a Feedback 15 This is analyzed by the Notification Generator (11). This generator analyzes the student's learning. a feedback signal based on its performance or the accuracy of its predictions This generated signal controls the parameters of the Corrective Module (3) in real time. It is passed back to the Corrective Module (3) to be updated. Thanks to this loop Corrective Module (3) adapts itself dynamically to the needs of the student and information 20 It continuously optimizes the correction process. This benefits the teacher, the corrector, and the student. An interactive and self-improving learning ecosystem is formed among them. Finally, the final results produced by the system are transmitted through an Output Unit (5). This output is presented to the user or subsequent systems. This output is only from the Student Model (4) Not only the classification accuracy, but also the successful return of the Corrective Module (3) 25 This can also include the amount of information gained. The model is an artificial intelligence model with an online correction mechanism to protect intellectual property. The intelligence learning method includes the following steps: 1001. Input Data Unit (1) by raw input data to teacher model (2) and Simultaneous transmission to the student model (4) 30 6 1002. Teacher Model (2) intentionally for the purpose of intellectual property protection generating perturbed forecasts 1003. Input Normalization Layer (6) located within Corrective Module (3) normalization of distorted estimates from the teacher model (2) 5. 1004. At least one Feature Correction Layer (7) located within Correction Module (3) by, perturbed information through nonlinear transformations correction and retrieval of meaningful information 1005. Corrected and reliable probability distributions by the Corrective Module (3). Generating the Corrected Prediction Output (8) and transferring it to the Student Model (4) 10 1006. By the Input Fusion Layer (9) located within the Student Model (4), Input Raw data from Data Unit (1) and corrected data from Corrective Module (3) combining predictions By the Prediction Layer (10) within the Student Model (4), 1007. Generating final class predictions based on fused data 15 1008. Feedback Generator (11) analyzes the outputs produced by the Student Model (4). by generating a feedback signal and sending this signal to the Corrective Module (3) transmission 1009. Output Unit (5) by Student Model (4) and / or Corrective Module (3) Presenting the final results from the source to the user or external systems. 20

Claims

7 REQUESTS 1. The model has an online correction mechanism to protect intellectual property. It is an artificial intelligence learning system whose characteristic is that it runs on a server. - by taking input data into the system, teacher model (2) and student model (4) at least one input data unit that enables transmission (1) 5 - Based on the input data, to make it difficult to copy from outside. Teacher model that produces deliberately distorted prediction outputs (2) - deliberately distorted prediction outputs from the teacher model (2) It has at least one component that makes it suitable for learning through processing. input normalization layer (6) and at least one feature correction layer (7) 10 containing, the hidden meaningful information within those distorted predictions, can be retrieved. structured and corrected probability distributions for winning producing at least one corrective module (3) - Corrective module (3) with raw input data provided from input data unit (1) The combined evaluation of the corrected forecast outputs produced by 15 providing, at least one input fusion layer (9) and at least one prediction based on the aforementioned fused information, which contains the layer (10), the final generating classroom predictions and correcting the learning process with the relevant teacher. At least one student model structured to optimize with knowledge (4) - Final class estimates and / or corrective 20 produced by the student model (4) Corrected teacher information recovered by module (3) is given to the user. or enabling the presentation of those outputs to external systems, in the form of digital data, will transmit visual output, reports, or machine-readable data in various formats at least one output unit structured in this way (5), - distorted prediction outputs and / or probability 25 from the teacher model (2) by organizing the statistical properties of their distributions, making them suitable for processing. bringing about these predictions with predetermined scaling, centering, with at least one of the reweighting or noise suppression processes normalizing and thus feature correction of normalized data. At least one input normalization configured to transmit to layer (7) 30 layer (6) - Prediction data normalized by input normalization layer (6) By working on it, the intentional stemming from the teacher model (2) 8 to suppress perturbations and the hidden meanings within those predictions structured, linear and / or designed to reveal information components At least one convolutional neural network that applies nonlinear transformations, complete the associated layer, attention mechanism, or a combination thereof, and At least one feature that enables the derivation of corrected probability distributions 5 correction layer (7), - by the corrective module (3), intentionally from the teacher model (2) through feature correction layers of distorted prediction outputs (7) obtained as a result of processing, with suppressed perturbation effects and latent meaning. 10 information that is recovered and can be directly used in the learning process at least one of the following forms of decision representation: probability distribution, score vector, or decision representation. a corrected forecast output (8) - Corrective module (3) with raw input data provided from input data unit (1) in a common representation space of corrected forecast outputs produced by enabling the combination and weighting of the data in question, 15 at least one of the following operations: merging, aligning, or feature transformation implementing and transmitting the fused features to the prediction layer (10) at least one input fusion layer structured to (9),  final based on fused features from the input fusion layer (9) linear and / or nonlinear 20 that generate class, score or regression estimates. transformations that do not involve and the probability distribution of the predictions in question, decision At least one prediction layer (10) that provides the score or label output in the form of an output. - analyze the outputs produced by the prediction layer (10) of the student model (4) by improving learning performance, prediction accuracy, or error metrics. generating at least one feedback signal based on and 25 of these signals cycle the parameters of the corrective module (3) and / or the input fusion layer (9) at least one feedback generator that sends feedback back to the system for internal updating. (11) It includes.

2. The model has an online correction mechanism to protect intellectual property. It is an artificial intelligence learning method, and its characteristic feature is; 9 1001. Input Data Unit (1) by raw input data to teacher model (2) and simultaneous transmission to the student model (4) 1002. Teacher Model (2) intentionally for the purpose of intellectual property protection. generating perturbed estimates 1003. Input Normalization Layer (6) located within Corrective Module (3) 5 normalize the distorted estimates from the teacher model (2) made ready for processing 1004. At least one Feature Correction Layer located within Correction Module (3) (7) by nonlinear transformations of perturbed information through correction and recovery of meaningful information 10 1005. Corrected and reliable probability by Corrective Module (3). Generating the Corrected Estimate Output (8) of the distributions and Student Transfer to the model (4) 1006. Input Fusion Layer (9) located within Student Model (4) by, raw data from Input Data Unit (1) and Corrective Module 15 (3) Combining the incoming corrected estimates By the Prediction Layer (10) within the Student Model (4), 1007. Generating final class predictions based on fused data. 1008. Outputs generated by the Student Model (4) by the Feedback Generator (11). by analyzing the feedback signal and applying Corrective 20 to this signal. Transmission to the module (3) 1009. Output Unit (5) by Student Model (4) and / or Corrective Module (3) presenting the final results from the source to the user or external systems. It includes the steps of the process.

3. Online revision to protect the model intellectual property in accordance with Claim 1 25 It is an artificial intelligence learning system with a mechanism; its characteristic is that it is a teacher. if the distorted estimates from the model (2) are used directly Corrective module (3) that prevents the learning process from being negatively affected It includes.

4. Online revision 30 to protect the model intellectual property in accordance with Claim 1. It is an artificial intelligence learning system with a linear mechanism; its characteristic feature is this. a convolutional neural network-based architecture to capture missing relationships and It includes a feature correction layer (7) containing activation functions.

5. Online revision to protect the model intellectual property in accordance with Claim 1. It is an artificial intelligence learning system with a mechanism that features input fusion. corrected teacher information with raw input data through layer (9) It includes a student model (4) that performs learning by integrating.

6. Online revision 5 to protect the model intellectual property in accordance with Claim 1. It is an artificial intelligence learning system with a mechanism whose characteristic is that the student both real data patterns and the teacher's refined knowledge It includes an input fusion layer (9) that enables it to learn simultaneously. 15 25

Images