Description:CROSS-REFERENCE TO RELATED APPLICATIONS AND PRIORITY
[0001] The present application does not claim priority from any other patent application.
TECHNICAL FIELD
[0002] The presently disclosed embodiments are related, in general, to a model training. More particularly, the presently disclosed embodiments are related to a method and a system for updating a model to classify authenticity.
BACKGROUND
[0003] This section is intended to introduce the reader to various aspects of art (the relevant technical field or area of knowledge to which the invention pertains), which may be related to various aspects of the present disclosure that are described or claimed below. This discussion is believed to be helpful in providing the reader with background information to facilitate a better understanding of the various aspects of the present disclosure. Accordingly, it should be understood that these statements in this background section are to be read in this light, and not as admissions of prior art. Similarly, a problem mentioned in the background section.
[0004] In the current digital content ecosystem, the amount and the quality of the synthetic content are increasing across the video content, the image content, the audio content, and the text content. Further, the number of the synthetic origins, including the generative models capable of producing the deepfake or synthetic content, continues to increase. This growth increases the operational need for the deepfake detection models that can detect the synthetic content and identify the origin of the synthetic content across the expanding set of the synthetic origins. The deepfake detection models require periodic updating to remain effective against the new training data associated with the one or more new synthetic origins. However, adapting the model to the new synthetic origins is not trivial, and typically involves time-consuming and resource-intensive retraining using the complete data or the combination of the old data and the new data. Further, such retraining can degrade the detection capabilities associated with the one or more synthetic origins previously learned by the model due to catastrophic forgetting, which reduces continuity of the detection performance.
[0005] Existing approaches that maintain multiple models, or multiple parts of the model, and dynamically select a branch using the mixture-of-experts techniques introduce additional complexity in the system design. Further, these approaches increase the compute overhead, the memory footprint, and the runtime latency, and complicate deployment, versioning, and maintenance across the model updates. Further, practical deployment environments often require updating the model under constrained resources, including limited compute budgets, limited memory, and limited access to the old training data due to storage constraints, privacy requirements, or licensing restrictions. As a result, the model update process can be delayed, and the deepfake detection model can remain outdated relative to the new synthetic origins, leading to reduced robustness against newly generated synthetic content.
[0006] Further, updating all portions of the model during adaptation can introduce unnecessary parameter changes that destabilize previously learned representations. Such full-parameter updates increase the risk of overfitting to the new training data and reduce the ability of the model to preserve the learned weight matrices associated with the earlier synthetic origins. This creates a need for training the model, while controlling the update impact on the corresponding learned weight matrices. The conventional update process becomes iterative and manual, increasing engineering effort and prolonging the time needed to deploy the updated model for identifying the origin of the synthetic content generated by the one or more new synthetic origins.
[0007] In view of the above, addressing the aforementioned technical challenges requires an improved method for updating a model.
[0008] Further limitations and disadvantages of conventional and traditional approaches will become apparent to one of skill in the art, through comparison of described systems with some aspects of the present disclosure, as set forth in the remainder of the present application and with reference to the drawings.
SUMMARY
[0009] This summary is provided to introduce concepts related to a method and a system for updating a model and the concepts are further described below in the detailed description. This summary is not intended to identify essential features of the claimed subject matter nor is it intended for use in determining or limiting the scope of the claimed subject matter.
[0010] According to embodiments illustrated herein, a computer implemented method for updating a model is disclosed. Further, the method may be implemented by a processor and a memory communicatively coupled to the processor, with the memory configured to store processor-executable programmed instructions. Further, the method may comprise a step of receiving a model trained using training data associated with one or more synthetic origins. Further, the method may comprise a step of receiving new training data associated with one or more new synthetic origins. Further, the method may comprise a step of selectively training one or more portions of the model using the new training data. The selection of the one or more portions of the model is performed based on generating rank-reduced approximations of corresponding learned weight matrices of the model. Furthermore, the method may comprise a step of updating the model to identify an origin of a synthetic content generated by the one or more new synthetic origins while maintaining detection capabilities of the model for the one or more synthetic origins.
[0011] According to embodiments illustrated herein, the system for updating the model is disclosed. Further, the system may comprise the processor and the memory communicatively coupled with the processor. Further, the memory may be configured to store the programmed instructions that cause the processor to perform the following operations. Further, the processor may be configured to receive the model trained using the training data associated with the one or more synthetic origins. Further, the processor may be configured to receive the new training data associated with the one or more new synthetic origins. Further, the processor may be configured to selectively train the one or more portions of the model utilizing the new training data. Further, selection of the one or more portions of the model is performed based on generating the rank-reduced approximations of the corresponding learned weight matrices of the model. Furthermore, the processor may be configured to update the model to identify the origin of the synthetic content generated by the one or more new synthetic origins while maintaining the detection capabilities of the model for the one or more synthetic origins.
[0012] According to embodiments illustrated herein, the non-transitory computer-readable storage medium having stored there on the set of computer-executable instructions for updating the model is disclosed. Further, the set of computer-executable instructions may be configured to cause the computer comprising the one or more processors to perform the steps. Further, the steps may comprise receiving the model trained using the training data associated with the one or more synthetic origins. Further, the steps may comprise receiving the new training data associated with the one or more new synthetic origins. Further, the steps may comprise selectively training the one or more portions of the model utilizing the new training data. The selection of the one or more portions of the model is performed based on generating the rank-reduced approximations of the corresponding learned weight matrices of the model. Furthermore, the steps may comprise updating the model to identify the origin of the synthetic content generated by the one or more new synthetic origins while maintaining the detection capabilities of the one or more synthetic origins.
[0013] The foregoing summary is illustrative only and is not intended to be in any way limiting. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features will become apparent by reference to the drawings and the following detailed description.
BRIEF DESCRIPTION OF DRAWINGS
[0014] The accompanying drawings illustrate the various embodiments of systems, methods, and other aspects of the disclosure. Any person with ordinary skills in art will appreciate that the illustrated element boundaries (e.g., boxes, groups of boxes, or other shapes) in the figures represent one example of the boundaries. In some examples, one element may be designed as multiple elements, or multiple elements may be designed as one element. In some examples, an element shown as an internal component of one element may be implemented as an external component in another, and vice versa. Further, the elements may not be drawn to scale.
[0015] Various embodiments will hereinafter be described in accordance with the appended drawings, which are provided to illustrate and not to limit the scope in any manner, and similar designations denote similar elements, and in which:
[0016] FIG. 1 is a block diagram that illustrates a system (100) for updating a model, in accordance with an embodiment of present subject matter.
[0017] FIG. 2 is a block diagram that illustrates various components of an application server (104) configured for performing steps of updating the model, in accordance with an embodiment of the present subject matter.
[0018] FIG. 3 is a flowchart that illustrates a method (300) for updating the model, in accordance with an embodiment of the present subject matter.
[0019] FIG. 4 illustrates a block diagram (400) of an exemplary computer system for implementing embodiments consistent with the present subject matter.
DETAILED DESCRIPTION
[0020] The present disclosure may be best understood with reference to the detailed figures and description set forth herein. Various embodiments are discussed below with reference to the figures. However, those skilled in the art will readily appreciate that the detailed descriptions given herein with respect to the figures are simply for explanatory purposes as the methods and systems may extend beyond the described embodiments. For example, the teachings presented, and the needs of a particular application may yield multiple alternative and suitable approaches to implement the functionality of any detail described herein. Therefore, any approach may extend beyond the particular implementation choices in the following embodiments described and shown.
[0021] References to “one embodiment,” “at least one embodiment,” “an embodiment,” “one example,” “an example,” “for example,” and so on indicate that the embodiment(s) or example(s) may include a particular feature, structure, characteristic, property, element, or limitation but that not every embodiment or example necessarily includes that particular feature, structure, characteristic, property, element, or limitation. Further, repeated use of the phrase “in an embodiment” does not necessarily refer to the same embodiment. The terms “comprise”, “comprising”, “include(s)”, or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a setup, system or method that comprises a list of components or steps does not include only those components or steps but may include other components or steps not expressly listed or inherent to such setup or system or method. In other words, one or more elements in a system or apparatus preceded by “comprises… a” does not, without more constraints, preclude the existence of other elements or additional elements in the system or apparatus.
[0022] An objective of the present disclosure is to update the model that is trained using the training data associated with the one or more synthetic origins, so that the model remains effective as the synthetic content generation landscape evolves.
[0023] Another objective of the present disclosure is to incorporate the new training data associated with the one or more new synthetic origins into the model with reduced time and resource consumption compared to retraining using the complete training data.
[0024] Yet another objective of the present disclosure is to determine the one or more portions of the model for adaptation based on the rank-reduced approximations of the corresponding learned weight matrices of the model, thereby limiting updates to targeted portions of the model.
[0025] Yet another objective of the present disclosure is to improve the ability of the model to identify the origin of the synthetic content generated by the one or more new synthetic origins after the model update.
[0026] Yet another objective of the present disclosure is to maintain the detection capabilities of the one or more synthetic origins supported by the model prior to incorporating the one or more new synthetic origins.
[0027] Yet another objective of the present disclosure is to reduce catastrophic forgetting associated with updating the model for the one or more new synthetic origins, thereby preserving learned representations for the one or more synthetic origins.
[0028] Yet another objective of the present disclosure is to reduce operational complexity associated with maintaining multiple models or multiple branches for different synthetic origins by enabling updates within the model.
[0029] FIG. 1 is a block diagram that illustrates a system (100) for updating the model, in accordance with an embodiment of present subject matter. The system (100) typically includes a database server (102), an application server (104), a communication network (106), and one or more portable devices (108). The database server (102), the application server (104), and the one or more portable devices (108) are typically communicatively coupled with each other via the communication network (106). In an embodiment, the application server (104) may communicate with the database server (102), and the one or more portable devices (108) using one or more protocols such as, but not limited to, Hypertext Transfer Protocol (HTTP), Transmission Control Protocol/Internet Protocol (TCP/IP)/User Datagram Protocol (UDP), Wireless Application Protocol (WAP), RF mesh, Bluetooth Low Energy (BLE), and the like, to communicate with one another.
[0030] In one embodiment, the database server (102) may refer to the computing device configured to store data related to the model, the training data associated with the one or more synthetic origins, and the new training data associated with the one or more new synthetic origins. Further, the database server (102) may store the learned weight matrices of the model and the rank-reduced approximations of the learned weight matrices. Further, the database server (102) may store data related to the adapter layers inserted into the base neural network, and the low-rank matrix decomposition components. Further, the database server (102) may store data related to the low-rank factorization, including the singular value decomposition outputs or the learned low-rank adapter parameters. Further, the database server (102) may store the reduction factor associated with the rank-reduced approximations, and data used for determining the reduction factor including the size of the new training data, the diversity of the new training data, and the complexity associated with the one or more new synthetic origins. Further, the database server (102) may store the semantic embeddings for the plurality of samples from the new training data and the vector database entries including the semantic embeddings and the associated metadata indicating the real classification labels or the fake classification labels. Further, the database server (102) may store the plurality of training batches, including the anchor sample, the positive sample, and the contrast sample, and the minimum embedding distance results used for retrieving the contrast sample. Further, the database server (102) may store the triplet loss function values computed using the anchor sample, the positive sample, and the contrast sample. Further, the database server (102) may store the plurality of clusters generated by clustering the semantic embeddings, and the centroid and the cluster variance for each of the plurality of clusters. Further, the database server (102) may store the predefined number of epochs for each of the one or more new synthetic origins.
[0031] In an embodiment, the database server (102) may include a special purpose operating system specifically configured to perform one or more database operations on the stored content. Examples of database operations may include, but are not limited to, Select, Insert, Update, and Delete. In an embodiment, the database server (102) may include hardware that may be configured to perform one or more predetermined operations. In an embodiment, the database server (102) may be realized through various technologies such as, but not limited to, Microsoft® SQL Server, Oracle®, IBM DB2®, Microsoft Access®, PostgreSQL®, DynamoDB®, MySQL®, SQLite®, MongoDB®, Cassandra®, Redis®, Azure®, Server distributed database technology and the like. In an embodiment, the database server (102) may be configured to utilize the application server (104) for updating the model. Further, the database server may be a vector database.
[0032] A person with ordinary skills in art will understand that the scope of the disclosure is not limited to the database server (102) as a separate entity. In an embodiment, the functionalities of the database server (102) can be integrated into the application server (104) or into the one or more portable devices (108).
[0033] In an embodiment, the application server (104) may refer to a computing device or a software framework hosting the application or a software service. In an embodiment, the application server (104) may be implemented to execute procedures such as, but not limited to, programs, policies, routines, or scripts stored in one or more memories for supporting the hosted application or the software service. In an embodiment, the hosted application or the software service may be configured to perform one or more predetermined operations. The application server (104) may be realized through various types of application servers such as, but are not limited to, a Java application server, a .NET framework application server, a Base4 application server, a PHP framework application server, or any other application server framework.
[0034] In an embodiment, the application server (104) may be configured to utilize the database server (102) for updating the model. In an implementation, the application server (104) may be configured to receive the model trained using the training data associated with the one or more synthetic origins. In an implementation, the application server (104) may be configured to receive the new training data associated with the one or more new synthetic origins. In one embodiment, the application server (104) may be configured to selectively train the one or more portions of the model utilizing the new training data, and select the one or more portions of the model based on generating the rank-reduced approximations of the corresponding learned weight matrices of the model. Further, the application server (104) may be configured to update the model to identify the origin of the synthetic content generated by the one or more new synthetic origins while maintaining the detection capabilities of the one or more synthetic origins.
[0035] In an embodiment, the communication network (106) may correspond to a communication medium through which the application server (104), the database server (102), and the one or more portable devices (108) may communicate with each other. Such a communication may be performed in accordance with various wired and wireless communication protocols. Examples of such wired and wireless communication protocols include, but are not limited to, Transmission Control Protocol and Internet Protocol (TCP/IP), User Datagram Protocol (UDP), Hypertext Transfer Protocol (HTTP), Wireless Application Protocol (WAP), File Transfer Protocol (FTP), ZigBee, EDGE, infrared IR), IEEE 802.11, 802.16, 2G, 3G, 4G, 5G, 6G, 7G cellular communication protocols, and/or Bluetooth (BT) communication protocols. The communication network (106) may either be a dedicated network or a shared network. Further, the communication network (106) may include a variety of network devices, including routers, bridges, servers, computing devices, storage devices, and the like. The communication network (106) may include, but is not limited to, the Internet, intranet, a cloud network, a Wireless Fidelity (Wi-Fi) network, a Wireless Local Area Network (WLAN), a Local Area Network (LAN), a cable network, the wireless network, a telephone network (e.g., Analog, Digital, POTS, PSTN, ISDN, xDSL), a telephone line (POTS), a Metropolitan Area Network (MAN), an electronic positioning network, an X.25 network, an optical network (e.g., PON), a satellite network (e.g., VSAT), a packet-switched network, a circuit-switched network, a public network, a private network, and/or other wired or wireless communications network configured to carry data.
[0036] In an embodiment, the one or more portable devices (108) may refer to a computing device used by a user. The one or more portable devices (108) may comprise of one or more processors and one or more memory. The one or more memories may include computer readable code that may be executable by one or more processors to perform predetermined operations. In an embodiment, the one or more portable devices (108) may present a web user interface for updating the model using the application server (104). Example web user interfaces presented on the one or more portable devices (108) to display information about the data. Examples of the one or more portable devices (108) may include, but are not limited to, a personal computer, a laptop, a computer desktop, a personal digital assistant (PDA), a mobile device, a tablet, or any other computing device.
[0037] The system (100) can be implemented using hardware, software, or a combination of both, which includes using where suitable, one or more computer programs, mobile applications, or “apps” by deploying either on-premises over the corresponding computing terminals or virtually over cloud infrastructure. The system (100) may include various micro-services or groups of independent computer programs which can act independently in collaboration with other micro-services. The system (100) may also interact with a third-party or external computer system. Internally, the system (100) may be the central processor of all for updating the model.
[0038] In one embodiment, the system (100) is configured for updating the model. The system (100) comprises the processor (202) (illustrated in Fig. 2) and the memory (204) (illustrated in Fig. 2) communicatively coupled with the processor (202). The memory (204) is configured to store the one or more executable instructions that, when executed by the processor (202), enable the system (100) to receive the model trained using the training data associated with the one or more synthetic origins. Further, the system (100) is configured to receive the new training data associated with the one or more new synthetic origins. Further, the system (100) is configured to selectively train the one or more portions of the model utilizing the new training data, and to select the one or more portions of the model based on generating the rank-reduced approximations of the corresponding learned weight matrices of the model. Furthermore, the system (100) is configured to update the model to identify the origin of the synthetic content generated by the one or more new synthetic origins while maintaining the detection capabilities of the one or more synthetic origins.
[0039] FIG. 2 illustrates a block diagram illustrating various components of the application server (104) configured for performing stepwise execution to update the model, in accordance with an embodiment of the present subject matter. Further, the FIG. 2 is explained in conjunction with the FIG. 1. Here the application server (104) preferably includes a processor (202), a memory (204), a transceiver (206), an Input/Output (208), a user interface unit (210), a receiving unit (212), a training unit (214), an updating unit (216), and a display unit (218). The processor (202) is further preferably communicatively coupled to the memory (204), the transceiver (206), the Input/Output unit (208), the user interface unit (210), the receiving unit (212), the training unit (214), the updating unit (216), and the display unit (218), while the transceiver (206) is preferably communicatively coupled to the communication network (106).
[0040] The processor (202) comprises suitable logic, circuitry, interfaces, and/or code that may be configured to execute a set of instructions stored in the memory (204), and may be implemented based on several processor technologies known in the art. The processor (202) works in coordination with the transceiver (206), the Input/Output unit (208), the user interface unit (210), the receiving unit (212), the training unit (214), the updating unit (216), and the display unit (218). Examples of the processor (202) include, but not limited to, standard microprocessor, microcontroller, central processing unit (CPU), an X86-based processor, a Reduced Instruction Set Computing (RISC) processor, an Application- Specific Integrated Circuit (ASIC) processor, and a Complex Instruction Set Computing (CISC) processor, distributed or cloud processing unit, state machines, logic circuitries, and/or any devices that manipulate signals based on operational instructions and/or other processing logic that accommodates the requirements of the present invention.
[0041] The memory (204) comprises suitable logic, circuitry, interfaces, and/or code that may be configured to store the set of instructions, which are executed by the processor (202). Preferably, the memory (204) is configured to store one or more programs, routines, or scripts that are executed in coordination with the processor (202). Additionally, the memory (204) may include any computer-readable medium or computer program product known in the art including, for example, volatile memory, such as static random-access memory (SRAM) and dynamic random-access memory (DRAM), and/or non-volatile memory, such as read only memory (ROM), erasable programmable ROM, a Hard Disk Drive (HDD), flash memories, Secure Digital (SD) card, Solid State Disks (SSD), optical disks, magnetic tapes, memory cards, virtual memory and distributed cloud storage. The memory (204) may be removable, non-removable, or a combination thereof. Further, the memory (204) may include routines, programs, objects, components, data structures, etc., which perform particular tasks or implement particular abstract data types. The memory (204) may include programs or coded instructions that supplement applications and functions of the system (100). In one embodiment, the memory (204), amongst other things, serves as a repository for storing data processed, received, and generated by one or more of the programs or the coded instructions. In yet another embodiment, the memory (204) may be managed under a federated structure that enables adaptability and responsiveness of the application server (104).
[0042] The transceiver (206) comprises suitable logic, circuitry, interfaces, and/or code that may be configured to receive, process or transmit information, data or signals, which are stored by the memory (204) and executed by the processor (202). The transceiver (206) is preferably configured to receive, process or transmit, one or more programs, routines, or scripts that are executed in coordination with the processor (202). The transceiver (206) is preferably communicatively coupled to the communication network (106) of the system (100) for communicating all the information, data, signal, programs, routines or scripts through the network. The transceiver (206) may be configured to receive a request for updating the model.
[0043] The transceiver (206) may implement one or more known technologies to support wired or wireless communication with the communication network (106). In an embodiment, the transceiver (206) may include, but is not limited to, an antenna, a radio frequency (RF) transceiver, one or more amplifiers, a tuner, one or more oscillators, a digital signal processor, a Universal Serial Bus (USB) device, a coder-decoder (CODEC) chipset, a subscriber identity module (SIM) card, and/or a local buffer. Also, the transceiver (206) may communicate via wireless communication with networks, such as the Internet, an Intranet and/or a wireless network, such as a cellular telephone network, a wireless local area network (LAN) and/or a metropolitan area network (MAN). Accordingly, the wireless communication may use any of a plurality of communication standards, protocols and technologies, such as: Global System for Mobile Communications (GSM), Enhanced Data GSM Environment (EDGE), wideband code division multiple access (W-CDMA), code division multiple access (CDMA), time division multiple access (TDMA), Bluetooth, Wireless Fidelity (Wi-Fi) (e.g., IEEE 802.11a, IEEE 802.11b, IEEE 802.11g and/or IEEE 802.11n), voice over Internet Protocol (VoIP), Wi-MAX, a protocol for email, instant messaging, and/or Short Message Service (SMS).
[0044] The input/output (I/O) unit (208) comprises suitable logic, circuitry, interfaces, and/or code that may be configured to receive or present information. The input/output unit (208) comprises various input and output devices that are configured to communicate with the processor (202). Examples of the input devices include, but are not limited to, a keyboard, a mouse, a joystick, a touch screen, a microphone, a camera, and/or a docking station. Examples of the output devices include, but are not limited to, a display screen and/or a speaker. The I/O unit (208) may include a variety of software and hardware interfaces, for example, a web interface, a graphical user interface, and the like. The I/O unit (208) may allow the system (100) to interact with the user directly or through the portable devices (108). Further, the I/O unit (208) may enable the system (100) to communicate with other computing devices, such as web servers and external data servers (not shown). The I/O unit (208) can facilitate multiple communications within a wide variety of networks and protocol types, including wired networks, for example, LAN, cable, etc., and wireless networks, such as WLAN, cellular, or satellite. The I/O unit (208) may include one or more ports for connecting a number of devices to one another or to another server. In one embodiment, the I/O unit (208) allows the application server (104) to be logically coupled to other portable devices (108), some of which may be built in. Illustrative components include tablets, mobile phones, desktop computers, wireless devices, etc.
[0045] In an embodiment, the user interface unit (210) of the application server (104) is disclosed. The user interface unit (210) comprises suitable logic, circuitry, interfaces, and/or code configured to provide an interactive interface for enabling user input and configuration of the model, the training data associated with the one or more synthetic origins, and the new training data associated with the one or more new synthetic origins. Further, the user interface unit (210) may enable user selection of the one or more new synthetic origins to be incorporated into the model, and may present outputs associated with identifying the origin of the synthetic content generated by the one or more synthetic origins and the one or more new synthetic origins.
[0046] In another embodiment, the receiving unit (212) of the system (100) for updating the model is disclosed. Further, the receiving unit (212) may be configured to receive the model trained using the training data associated with the one or more synthetic origins. Further, the receiving unit (212) may be configured to receive the new training data associated with the one or more new synthetic origins. Further, the model may correspond to the authenticity classification model configured to distinguish between the genuine content and the synthetic content in the one or more modalities selected from the video, the image, the audio, and the text. Further, the model may be configured to identify the origin of the synthetic content from the one or more synthetic origins. Furthermore, the new training data may correspond to the plurality of new deepfake generation origins for which the prior deepfake detection training data was unavailable.
[0047] In another embodiment, the training unit (214) of the system (100) for updating the model is disclosed. Further, the training unit (214) may be configured to selectively train the one or more portions of the model utilizing the new training data. Further, the training unit (214) may be configured to select the one or more portions of the model based on generating the rank-reduced approximations of the corresponding learned weight matrices of the model. Further, the one or more portions of the model may comprise the adapter layers inserted into the base neural network, and the adapter layers may comprise the low-rank matrix decomposition components. Further, the training unit (214) may be configured to generate the rank-reduced approximations using the low-rank factorization of the learned weight matrices, and the low-rank factorization may correspond to at least one of the singular value decomposition (SVD) or the learned low-rank adapters (LoRA).
[0048] Further, the training unit (214) may be configured to associate the reduction factor with the rank-reduced approximations. Further, the training unit (214) may be configured to determine the reduction factor based on at least one of the size of the new training data, the diversity of the new training data, the complexity associated with the one or more new synthetic origins, or the combination thereof. Further, the reduction factor may be inversely proportional to the size of the new training data associated with the one or more new synthetic origins. Further, the training unit (214) may be configured to dynamically update the reduction factor during training based on the performance metrics.
[0049] Further, the training unit (214) may be configured to perform the selectively training of the one or more portions of the model using the plurality of training batches corresponding to the new training data. Further, the training unit (214) may be configured to generate the plurality of training batches by generating the semantic embeddings for the plurality of samples from the new training data and storing the semantic embeddings in the vector database. Further, the vector database may be configured to store the embeddings and the associated metadata indicating the real classification labels or the fake classification labels. Further, the training unit (214) may be configured to identify the positive sample from the plurality of samples associated with the same class as the anchor sample. Further, the training unit (214) may be configured to retrieve the contrast sample from the plurality of samples from the vector database with the opposite class with the minimum embedding distance from the anchor sample by performing the nearest-neighbour search in the vector database subject to an opposite class constraint. Further, the batching of data may be performed to ensure maximum information exposure to the model.
[0050] Furthermore, the training unit (214) may be configured to perform the selectively training of the one or more portions of the model using the plurality of training batches based at least in part on the triplet loss function computed using the anchor sample, the positive sample, and the contrast sample. Further, the triplet loss function may enforce the margin between the embeddings of the anchor sample and the contrast sample. Further, the positive sample may be selected from the vector database based on the minimum embedding distance within the same class.
[0051] Further, the training unit (214) may be configured to cluster the semantic embeddings associated with the new training data into the plurality of clusters corresponding to the number of the one or more new synthetic origins. Further, the training unit (214) may be configured to compute the centroid and the cluster variance of each of the plurality of clusters. Further, the training unit (214) may be configured to generate the plurality of training batches based on the distances from the cluster centroids. Further, the training unit (214) may be configured to perform the selectively training of the one or more portions of the model for the predefined number of epochs for each of the one or more new synthetic origins. Further, the training unit (214) may be configured to perform the selective training to prevent catastrophic forgetting of the previously learned deepfake or synthetic origin sources. Further, the training unit (214) may be configured to exclude retraining of the core base model parameters. In an exemplary embodiment, the training unit (214) may further comprise computing the orthogonal embedding loss to enforce orthogonality between the embeddings associated with different deepfake generation origins.
[0052] In one embodiment, the updating unit (216) of the system (100) is disclosed. The updating unit (216) may be configured to update the model to identify the origin of the synthetic content generated by the one or more new synthetic origins, while maintaining the detection capabilities of the one or more synthetic origins. Further, the updating unit (216) may be configured to incorporate the one or more new synthetic origins into the one or more synthetic origins corresponding to the model. Furthermore, the updating unit (216) may be configured to perform the update without degrading the detection performance for the previously learned origins.
[0053] Further, the display unit (218) may be configured for generating and presenting information related to updating the model to identify the origin of the synthetic content generated by the one or more new synthetic origins while maintaining the detection capabilities of the one or more synthetic origins. The displayed information may comprise at least one of the model update status, the incorporation status of the one or more new synthetic origins into the one or more synthetic origins corresponding to the model, and outputs associated with the origin identification performed by the updated model. In an embodiment, the display unit (218) may be configured to present information related to the rank-reduced approximations of the corresponding learned weight matrices of the model, and information related to the reduction factor used during selectively training of the one or more portions of the model. Further, the display unit (218) may be configured to present information related to the plurality of training batches generated from the new training data, including information related to the anchor sample, the positive sample, and the contrast sample, and information related to the minimum embedding distance used for retrieving the contrast sample from the vector database. Furthermore, the display unit (218) may be configured to present information related to the triplet loss function computed using the anchor sample, the positive sample, and the contrast sample. In another embodiment, the display unit (218) may be configured to present information related to clustering of the semantic embeddings associated with the new training data, including the plurality of clusters, the centroid, and the cluster variance of each of the plurality of clusters, and information indicating that the selectively training is performed for the predefined number of epochs for each of the one or more new synthetic origins.
[0054] In an embodiment, the application server (104) of the system (100) is configured for identifying the origin of the synthetic content generated by the one or more new synthetic origins based on utilizing at least one of the updated model, class information derived from the embeddings and metadata stored in the vector database, or a combination thereof. It is important to note that identification of the origin of the synthetic content by utilizing both the updated model and the class information derived from the vector database, is to cross check the competitive quality of the output generated by the updated model, so that a corrective measure can be take for improving output of the updated model.
[0055] A person skilled in the art will understand that the scope of the disclosure should not be limited to a single domain and using the aforementioned techniques. Further, the examples provided in supra are for illustrative purposes and should not be construed to limit the scope of the disclosure.
[0056] Referring to Fig. 3, a flowchart that illustrates a method (300) for updating the model, in accordance with at least one embodiment of the present subject matter. The method (300) may be implemented by the application server (104) including the processor (202) and the memory (204) communicatively coupled to the processor (202) and the memory (204) is configured to store processor-executable programmed instructions, caused the processor (202) to perform the following steps.
[0057] At step (302), the processor (202) may be configured to receive the model, and the model is trained using the training data associated with the one or more synthetic origins.
[0058] At step (304), the processor (202) may be configured to receive the new training data associated with the one or more new synthetic origins.
[0059] At step (306), the processor (202) may be configured to selectively train the one or more portions of the model utilizing the new training data. Further, the one or more portions of the model is selected based on generating rank-reduced approximations of corresponding learned weight matrices of the model.
[0060] At step (308), the processor (202) may be configured to update the model to identify an origin of a synthetic content generated by the one or more new synthetic origins while maintaining detection capabilities of the one or more synthetic origins.
[0061] Let us delve into a detailed working example of the present disclosure.
[0062] Working Example 1:
[0063] Consider the model trained using the training data associated with the one or more synthetic origins S1 and S2. The model corresponds to the authenticity classification model configured to distinguish between the genuine content and the synthetic content in the video modality, the image modality, the audio modality, and the text modality, and the model is configured to identify the origin of the synthetic content as S1 or S2. After deployment, the new training data associated with the one or more new synthetic origins S3 is received, and S3 corresponds to the new synthetic origin not represented in the training data used to train the model.
[0064] The method (300) generates the rank-reduced approximations of the corresponding learned weight matrices of the model using the low-rank factorization that corresponds to the singular value decomposition (SVD) or the learned low-rank adapters (LoRA). Based on the rank-reduced approximations, the method (300) selects the one or more portions of the model for selective training, and the selected portions include the adapter layers inserted into the base neural network that include the low-rank matrix decomposition components. The reduction factor associated with the rank-reduced approximations is determined based on the size of the new training data, the diversity of the new training data, and the complexity associated with S3, and the reduction factor is inversely proportional to the size of the new training data associated with S3.
[0065] The method (300) generates the plurality of training batches from the new training data by generating the semantic embeddings for the plurality of samples and storing the semantic embeddings in the vector database along with the metadata indicating the real classification labels or the fake classification labels. For an anchor sample labeled fake, a positive sample is identified with the same class, and a contrast sample is retrieved from the vector database with the opposite class and the minimum embedding distance from the anchor sample.
[0066] The selective training is performed using the plurality of training batches based at least in part on the triplet loss function computed using the anchor sample, the positive sample, and the contrast sample, and the positive sample is selected from the vector database based on the minimum embedding distance within the same class. The semantic embeddings associated with the new training data are clustered into the plurality of clusters corresponding to the number of the one or more new synthetic origins, and the centroid and the cluster variance are computed for each cluster. The plurality of training batches is generated based on the distances from the cluster centroids, and the selectively training is performed for the predefined number of epochs for S3.
[0067] After completing the selective training, the method (300) updates the model by incorporating S3 into the one or more synthetic origins corresponding to the model. The updated model identifies the origin of the synthetic content generated by S3 while maintaining the detection capabilities of S1 and S2, so that detection performance for the previously learned sources remains substantially unchanged.
[0068] Working Example 2:
[0069] Consider the model trained using the training data associated with the synthetic origins V1 and V2, and deployed in a real-time video screening pipeline. The model corresponds to the authenticity classification model configured to distinguish between the genuine video content and the synthetic video content, and the model is configured to identify the origin of the synthetic video content as V1 or V2. During operation, the new training data associated with the new synthetic origin V3 is received, and V3 corresponds to the new video deepfake generation source that begins producing synthetic videos in real time.
[0070] The method (300) generates the rank-reduced approximations of the corresponding learned weight matrices of the model using the low-rank factorization that corresponds to the singular value decomposition (SVD) or the learned low-rank adapters (LoRA). Based on the rank-reduced approximations, the method (300) selects the one or more portions of the model for selective training, and the selected portions comprise the adapter layers inserted into the base neural network that comprise the low-rank matrix decomposition components. The reduction factor associated with the rank-reduced approximations is determined based on the size of the new training data, the diversity of the new training data, and the complexity associated with V3, and the reduction factor is inversely proportional to the size of the new training data associated with V3.
[0071] The method (300) generates the plurality of training batches corresponding to the new training data by generating the semantic embeddings for the plurality of video samples and storing the semantic embeddings in the vector database along with the associated metadata indicating the real classification labels or the fake classification labels. For an anchor video sample labeled fake, a positive sample is identified with the same class, and a contrast sample is retrieved from the vector database with the opposite class and the minimum embedding distance from the anchor sample.
[0072] The selective training of the selected portions of the model is performed using the plurality of training batches based at least in part on the triplet loss function computed using the anchor sample, the positive sample, and the contrast sample, and the positive sample is selected from the vector database based on the minimum embedding distance within the same class. The semantic embeddings associated with the new training data are clustered into the plurality of clusters corresponding to the number of the new synthetic origins, and the centroid and the cluster variance are computed for each cluster. The plurality of training batches is generated based on distances from the cluster centroids, and the selectively training is performed for the predefined number of epochs for V3.
[0073] After completing the selectively training, the method (300) updates the model by incorporating V3 into the synthetic origins corresponding to the model. The updated model identifies the origin of the synthetic video content generated by V3 while maintaining the detection capabilities of V1 and V2, and the deployed video screening pipeline continues to classify the genuine video content and the synthetic video content while assigning origin labels across V1, V2, and V3.
[0074] A person skilled in the art will understand that the scope of the disclosure is not limited to scenarios based on the aforementioned factors and using the aforementioned techniques and that the examples provided do not limit the scope of the disclosure.
[0075] A person skilled in the art will understand that the scope of the disclosure is not limited to scenarios based on the aforementioned factors and using the aforementioned techniques, and that the examples provided do not limit the scope of the disclosure.
[0076] FIG. 4 illustrates a block diagram of an exemplary computer system (401) for implementing embodiments consistent with the present disclosure.
[0077] Variations of computer system (401) may be used for updating the model. The computer system (401) may comprise a central processing unit (“CPU” or “processor”) (402). The processor (402) may comprise at least one data processor for executing program components for executing user or system generated requests. A user may include a person, a person using a device such as those included in this disclosure, or such a device itself. Additionally, the processor (402) may include specialized processing units such as integrated system (bus) controllers, memory management control units, floating point units, graphics processing units, digital signal processing units, or the like. In various implementations the processor (402) may include a microprocessor, such as AMD Athlon, Duron or Opteron, ARM’s application, embedded or secure processors, IBM PowerPC, Intel’s Core, Itanium, Xeon, Celeron or other line of processors, for example. Accordingly, the processor (402) may be implemented using mainframe, distributed processor, multi-core, parallel, grid, or other architectures. Some embodiments may utilize embedded technologies like application-specific integrated circuits (ASICs), digital signal processors (DSPs), or Field Programmable Gate Arrays (FPGAs), for example.
[0078] Processor (402) may be disposed in communication with one or more input/output (I/O) devices via I/O interface (403). Accordingly, the I/O interface (403) may employ communication protocols/methods such as, without limitation, audio, analog, digital, monoaural, RCA, stereo, IEEE-1394, serial bus, universal serial bus (USB), infrared, PS/2, BNC, coaxial, component, composite, digital visual interface (DVI), high-definition multimedia interface (HDMI), RF antennas, S-Video, VGA, IEEE 802.n /b/g/n/x, Bluetooth, cellular (e.g., code-division multiple access (CDMA), high-speed packet access (HSPA+), global system for mobile communications (GSM), long-term evolution (LTE), WiMAX, or the like, for example.
[0079] Using the I/O interface (403), the computer system (401) may communicate with one or more I/O devices. For example, the input device (404) may be an antenna, keyboard, mouse, joystick, (infrared) remote control, camera, card reader, fax machine, dongle, biometric reader, microphone, touch screen, touchpad, trackball, sensor (e.g., accelerometer, light sensor, GPS, gyroscope, proximity sensor, or the like), stylus, scanner, storage device, transceiver, video device/source, or visors, for example. Likewise, an output device (405) may be a user’s smartphone, tablet, cell phone, laptop, printer, computer desktop, fax machine, video display (e.g., cathode ray tube (CRT), liquid crystal display (LCD), light- emitting diode (LED), plasma, or the like), or audio speaker, for example. In some embodiments, a transceiver (406) may be disposed in connection with the processor (402). The transceiver (406) may facilitate various types of wireless transmission or reception. For example, the transceiver (406) may include an antenna operatively connected to a transceiver chip (example devices include the Texas Instruments® WiLink WL1283, Broadcom® BCM4750IUB8, Infineon Technologies® X-Gold 618-PMB9800, or the like), providing IEEE 802.11a/b/g/n, Bluetooth, FM, global positioning system (GPS), and/or 2G/3G/5G/6G HSDPA/HSUPA communications, for example.
[0080] In some embodiments, the processor (402) may be disposed in communication with a communication network (408) via a network interface (407). The network interface (407) is adapted to communicate with the communication network (408). The network interface (407) may employ connection protocols including, without limitation, direct connect, Ethernet (e.g., twisted pair 10/100/1000 Base T), transmission control protocol/internet protocol (TCP/IP), token ring, or IEEE 802.11a/b/g/n/x, for example. The communication network (408) may include, without limitation, a direct interconnection, local area network (LAN), wide area network (WAN), wireless network (e.g., using Wireless Application Protocol), or the Internet, for example. Using the network interface (407) and the communication network (408), the computer system (401) may communicate with devices such as shown as a laptop (409) or a mobile/cellular phone (410). Other exemplary devices may include, without limitation, personal computer(s), server(s), fax machines, printers, scanners, various mobile devices such as cellular telephones, smartphones (e.g., Apple iPhone, Blackberry, Android-based phones, etc.), tablet computers, desktop computers, eBook readers (Amazon Kindle, Nook, etc.), laptop computers, notebooks, gaming consoles (Microsoft Xbox, Nintendo DS, Sony PlayStation, etc.), or the like. In some embodiments, the computer system (401) may itself embody one or more of these devices.
[0081] In some embodiments, the processor (402) may be disposed in communication with one or more memory devices (e.g., RAM 413, ROM 414, etc.) via a storage interface (412). The storage interface (412) may connect to memory devices including, without limitation, memory drives, removable disc drives, etc., employing connection protocols such as serial advanced technology attachment (SATA), integrated drive electronics (IDE), IEEE-1394, universal serial bus (USB), fiber channel, small computer systems interface (SCSI), etc. The memory drives may further include a drum, magnetic disc drive, magneto-optical drive, optical drive, redundant array of independent discs (RAID), solid-state memory devices, or solid-state drives, for example.
[0082] The memory devices may store a collection of program or database components, including, without limitation, an operating system (416), user interface application (417), web browser (418), mail client/server (419), user/application data (420) (e.g., any data variables or data records discussed in this disclosure) for example. The operating system (416) may facilitate resource management and operation of the computer system (401). Examples of operating systems include, without limitation, Apple Macintosh OS X, UNIX, Unix-like system distributions (e.g., Berkeley Software Distribution (BSD), FreeBSD, NetBSD, OpenBSD, etc.), Linux distributions (e.g., Red Hat, Ubuntu, Kubuntu, etc.), IBM OS/2, Microsoft Windows (XP, Vista/7/8, etc.), Apple iOS, Google Android, Blackberry OS, or the like.
[0083] The user interface (417) is for facilitating the display, execution, interaction, manipulation, or operation of program components through textual or graphical facilities. For example, user interfaces (417) may provide computer interaction interface elements on a display system operatively connected to the computer system (401), such as cursors, icons, check boxes, menus, scrollers, windows, or widgets, for example. Graphical user interfaces (GUIs) may be employed, including, without limitation, Apple Macintosh operating systems’ Aqua, IBM OS/2, Microsoft Windows (e.g., Aero, Metro, etc.), Unix X-Windows, or web interface libraries (e.g., ActiveX, Java, JavaScript, AJAX, HTML, Adobe Flash, etc.), for example.
[0084] In some embodiments, the computer system (401) may implement a web browser (418) stored program component. The web browser (418) may be a hypertext viewing application, such as Microsoft Internet Explorer, Google Chrome, Mozilla Firefox, Apple Safari, or Microsoft Edge, for example. Secure web browsing may be provided using HTTPS (secure hypertext transport protocol), secure sockets layer (SSL), Transport Layer Security (TLS), or the like. Web browsers may utilize facilities such as AJAX, DHTML, Adobe Flash, JavaScript, Java, or application programming interfaces (APIs), for example. In some embodiments the computer system (401) may implement a mail client/server (419) stored program component. The mail server (419) may be an Internet mail server such as Microsoft Exchange, or the like. The mail server may utilize facilities such as ASP, ActiveX, ANSI C++/C#, Microsoft .NET, CGI scripts, Java, JavaScript, PERL, PHP, Python, or WebObjects, for example. The mail server (419) may utilize communication protocols such as internet message access protocol (IMAP), messaging application programming interface (MAPI), Microsoft Exchange, post office protocol (POP), simple mail transfer protocol (SMTP), or the like. In some embodiments, the computer system (401) may implement a mail client (420) stored program component. The mail client (420) may be a mail viewing application, such as Apple Mail, Microsoft Entourage, Microsoft Outlook, or Mozilla Thunderbird.
[0085] In some embodiments, the computer system (401) may store user/application data (421), such as the data, variables, records, or the like as described in this disclosure. Such databases may be implemented as fault-tolerant, relational, scalable, secure databases such as Oracle or Sybase, for example. Alternatively, such databases may be implemented using standardized data structures, such as an array, hash, linked list, struct, structured text file (e.g., JSON, XML), table, or as object-oriented databases (e.g., using ObjectStore, Poet, Zope, etc.). Such databases may be consolidated or distributed, sometimes among the various computer systems discussed above in this disclosure. It is to be understood that the structure and operation of any computer or database component may be combined, consolidated, or distributed in any working combination.
[0086] Furthermore, one or more computer-readable storage media may be utilized in implementing embodiments consistent with the present invention. A computer-readable storage medium refers to any type of physical memory on which information or data readable by a processor may be stored. Thus, a computer-readable storage medium may store instructions for execution by one or more processors, including instructions for causing the processor(s) to perform steps or stages consistent with the embodiments described herein. The term “computer readable medium” should be understood to include tangible items and exclude carrier waves and transient signals, i.e., non-transitory. Examples include Random Access Memory (RAM), Read- Only Memory (ROM), volatile memory, non-volatile memory, hard drives, Compact Disc (CD) ROMs, Digital Video Disc (DVDs), flash drives, disks, and any other known physical storage media.
[0087] The present disclosure addresses limitations in maintaining authenticity classification models in the current digital content ecosystem, in which the amount and the quality of the synthetic content are increasing across the video content, the image content, the audio content, and the text content. Further, the number of the synthetic origins capable of producing the deepfake or synthetic content continues to increase, which increases the operational need for authenticity classification models that can detect the synthetic content and identify the origin of the synthetic content across the expanding set of the synthetic origins. However, periodic updating of such models to remain effective against the new training data associated with the one or more new synthetic origins is time-consuming and resource-intensive, and retraining using the complete data or the combination of the old data and the new data can degrade detection capabilities for previously learned synthetic origins due to catastrophic forgetting. Further, multi-model approaches, including mixture-of-experts, increase system complexity, compute overhead, memory footprint, runtime latency, and maintenance burden, and constrained deployment environments can delay updates due to limited compute, limited memory, and limited access to the old training data.
[0088] Further, to address these challenges, the disclosed system (100) and method (300) provide an approach for updating the model trained using the training data associated with the one or more synthetic origins by receiving the model and receiving the new training data associated with the one or more new synthetic origins. The method (300) selectively trains the one or more portions of the model utilizing the new training data, and selects the one or more portions based on generating the rank-reduced approximations of the corresponding learned weight matrices of the model. Further, the method (300) updates the model to identify the origin of the synthetic content generated by the one or more new synthetic origins while maintaining the detection capabilities of the one or more synthetic origins.
[0089] Further, the disclosed system (100) and method (300) provide advantages over conventional update techniques by enabling faster adaptation to the one or more new synthetic origins while reducing reliance on full retraining using the complete training data. By limiting training to the one or more portions of the model selected based on the rank-reduced approximations of the corresponding learned weight matrices, the method (300) improves update efficiency and enables parameter-efficient adaptation through rank-reduced updates. Further, the method (300) supports incorporation of the one or more new synthetic origins into the model while maintaining detection capabilities of the one or more synthetic origins, thereby reducing catastrophic forgetting and avoiding degradation in detection performance for previously learned origins. Further, the method (300) supports broad applicability across modalities including the video modality, the image modality, the audio modality, and the text modality.
[0090] Various embodiments of the disclosure encompass numerous advantages, including methods and systems for updating the model. The disclosed method (300) and system (100) have several technical advantages, including but not limited to the following:
• Faster adaptation to the one or more new synthetic origins: The method (300) updates the model using the new training data associated with the one or more new synthetic origins, reducing reliance on full retraining using the complete training data.

• Selective training of the one or more portions of the model: The method (300) limits training to the one or more portions of the model selected based on the rank-reduced approximations of the corresponding learned weight matrices of the model, improving update efficiency.

• Rank-reduced update mechanism: By generating the rank-reduced approximations using the low-rank factorization corresponding to the singular value decomposition (SVD) or the learned low-rank adapters (LoRA), the method (300) enables parameter-efficient model adaptation.

• Reduced catastrophic forgetting: The method (300) updates the model to incorporate the one or more new synthetic origins while maintaining the detection capabilities of the one or more synthetic origins, reducing degradation in detection performance for previously learned sources.

• Adapter-layer based extensibility: The one or more portions of the model may comprise the adapter layers inserted into the base neural network with the low-rank matrix decomposition components, allowing updates without retraining the entire base neural network.

• Data-efficient batch formation using the vector database: The method (300) generates semantic embeddings, stores the semantic embeddings with the metadata indicating the real classification labels or the fake classification labels, and forms training batches by selecting the positive sample and retrieving the contrast sample with the minimum embedding distance, improving training signal quality from the new training data.

• Improved separation using the triplet loss function: The selectively training is performed based at least in part on the triplet loss function computed using the anchor sample, the positive sample, and the contrast sample, which enforces separation between the embeddings of different classes.

• Origin-aware organization of the new training data: By clustering the semantic embeddings into the plurality of clusters corresponding to the number of the one or more new synthetic origins and generating training batches based on distances from the cluster centroids, the method (300) supports structured adaptation to the one or more new synthetic origins.

• Controlled training duration per new synthetic origin: The selectively training is performed for the predefined number of epochs for each of the one or more new synthetic origins, enabling repeatable and bounded update cycles.

• Multi-modality coverage: The model supports distinguishing between the genuine content and the synthetic content in the one or more modalities selected from the video modality, the image modality, the audio modality, and the text modality, enabling broad applicability across synthetic content types.
[0091] In summary, the method (300) addresses limitations of existing deepfake detection update approaches by enabling efficient incorporation of the one or more new synthetic origins into the model using the new training data, without requiring retraining on the complete training data. By selectively training the one or more portions of the model identified through rank-reduced approximations of the corresponding learned weight matrices, and by leveraging low-rank factorization techniques including singular value decomposition (SVD) or learned low-rank adapters (LoRA), the method (300) improves update efficiency while reducing disruption to previously learned detection behavior. Further, the method (300) forms the plurality of training batches using semantic embeddings stored in the vector database, applies the triplet loss function, and optionally uses clustering-based batch construction, thereby strengthening training signals from the new training data. As a result, the updated model identifies the origin of the synthetic content generated by the one or more new synthetic origins while maintaining detection capabilities of the one or more synthetic origins, supporting robust multi-modality authenticity classification across video, image, audio, and text.
[0092] The claimed invention of a system (100) and a method (300) for updating the model involves tangible components, processes, and functionalities that interact to achieve specific technical outcomes. The system (100) integrates various elements such as processors, memory, databases, modelling, real-time fast processing, unnecessary data omitting and informed display techniques to effectively execute update of the model.
[0093] The present disclosure introduces a non-trivial combination of technologies and methodologies that provide a technical solution for a technical problem. While individual components like processors, databases, encryption, authorization and authentication are well-known in the field of computer science, their integration into a comprehensive system (100) for updating the model brings about an improvement and technical advancement in the field of authenticity classification.
[0094] In light of the above-mentioned advantages and the technical advancements provided by the disclosed method (300) and system (100) for updating the model, the claimed steps as discussed above are not routine, conventional, or well understood in the art, as the claimed steps enable the following solutions to the existing problems in conventional technologies. Further, the claimed steps clearly bring an improvement in the functioning of the device itself as the claimed steps provide a technical solution to a technical problem.
[0095] The present disclosure may be realized in hardware, or a combination of hardware and software. The present disclosure may be realized in a centralized fashion, in at least one computer system, or in a distributed fashion, where different elements may be spread across several interconnected computer systems. A computer system or other apparatus adapted for carrying out the methods described herein may be suited. A combination of hardware and software may be a general-purpose computer system with a computer program that, when loaded and executed, may control the computer system such that it carries out the methods described herein. The present disclosure may be realized in hardware that comprises a portion of an integrated circuit that also performs other functions.
[0096] A person with ordinary skills in the art will appreciate that the systems, modules, and sub-modules have been illustrated and explained to serve as examples and should not be considered limiting in any manner. It will be further appreciated that the variants of the above disclosed system elements, modules, and other features and functions, or alternatives thereof, may be combined to create other different systems or applications.
[0097] Those skilled in the art will appreciate that any of the aforementioned steps and/or system modules may be suitably replaced, reordered, or removed, and additional steps and/or system modules may be inserted, depending on the needs of a particular application. In addition, the systems of the aforementioned embodiments may be implemented using a wide variety of suitable processes and system modules, and are not limited to any particular computer hardware, software, middleware, firmware, microcode, and the like. The claims can encompass embodiments for hardware and software, or a combination thereof.
[0098] While the present disclosure has been described with reference to certain embodiments, it will be understood by those skilled in the art that various changes may be made, and equivalents may be substituted without departing from the scope of the present disclosure. In addition, many modifications may be made to adapt a particular situation or material to the teachings of the present disclosure without departing from its scope. Therefore, it is intended that the present disclosure is not limited to the particular embodiment disclosed, but that the present disclosure will include all embodiments falling within the scope of the appended claims. , Claims:WE CLAIM:
1. A computer implemented method (300) for updating a model, wherein the method (300) comprises:
receiving (302), via a processor (202), a model, wherein the model is trained using training data associated with one or more synthetic origins;
receiving (304), via the processor (202), a new training data associated with one or more new synthetic origins;
selectively training (306), via the processor (202), one or more portions of the model utilizing the new training data, wherein the one or more portions of the model is selected based on generating rank-reduced approximations of corresponding learned weight matrices of the model; and
updating (308), via the processor (202), the model to identify an origin of a synthetic content generated by the one or more new synthetic origins while maintaining detection capabilities of the model for the one or more synthetic origins.

2. The computer implemented method (300) as claimed in claim 1,
wherein the model corresponds to an authenticity classification model configured to distinguish between genuine and synthetic content in one or more modalities selected from video, image, audio, and text;
wherein the model is configured to identify an origin of the synthetic content from the one or more synthetic origins; and
wherein updating the model corresponds to incorporating the one or more new synthetic

3. The computer implemented method (300) as claimed in claim 1, wherein the one or more portions of the model comprise adapter layers inserted into a base neural network, wherein the adapter layers comprise low-rank matrix decomposition components.

4. The computer implemented method (300) as claimed in claim 1, wherein the rank-reduced approximations are generated using a low-rank factorization of the learned weight matrices, wherein the low-rank factorization corresponds to at least one of singular value decomposition (SVD) or learned low-rank adapters (LoRA).

5. The computer implemented method (300) as claimed in claim 1, wherein a reduction factor is associated with the rank-reduced approximations, wherein the reduction factor is determined based on at least one of a size of the new training data, a diversity of the new training data, complexity associated with the one or more new synthetic origins, or a combination thereof; wherein the reduction factor is inversely proportional to the size of new training data associated with the one or more new synthetic origins.

6. The computer implemented method (300) as claimed in claim 1, wherein the selectively training of the one or more portions of the model is performed a plurality of training batches corresponding to the new training data, wherein the plurality of training batches is generated by:
generating semantic embeddings for a plurality of samples from the new training data;
storing the semantic embeddings in a vector database, wherein the vector database is configured to store embeddings and associated metadata indicating real or fake classification labels;
identifying a positive sample from the plurality of samples, associated with a same class as an anchor sample; and
retrieving a contrast sample from the plurality of samples, from the vector database with an opposite class with a minimum embedding distance from the anchor sample.
7. The computer implemented method (300) as claimed in claim 6, wherein the selectively training of the one or more portions of the model using the plurality of training batches is performed based at least in part on a triplet loss function computed using the anchor sample, the positive sample, and the contrast sample; wherein the positive sample is selected from the vector database based on minimum embedding distance within a same class.

8. The computer implemented method (300) as claimed in claim 6, comprises identifying the origin of the synthetic content generated by the one or more new synthetic origins based on utilizing at least one of the updated model, class information derived from the embeddings and metadata stored in the vector database, or a combination thereof.

9. The computer implemented method (300) as claimed in claim 6, comprises
clustering the semantic embeddings associated with the new training data into a plurality of clusters corresponding to the number of one or more new synthetic origins; and
computing centroid and cluster variance of each of the plurality of clusters;
wherein generating of the plurality of training batches is based on distances from the cluster centroids.

10. The computer implemented method (300) as claimed in claim 1, wherein the selectively training of the one or more portions of the model is performed for a predefined number of epochs for each of the one or more new synthetic origins.

11. A system (100) for updating a model, the system (100) comprises:
a processor (202), a memory (204) communicatively coupled with the processor (202), wherein the memory (204) is configured to store one or more executable instructions, which cause the processor (202) to:
receive (302), a model, wherein the model is trained using training data associated with one or more synthetic origins;
receive (304), a new training data associated with one or more new synthetic origins;
selectively train (306), one or more portions of the model utilizing the new training data, wherein the one or more portions of the model is selected based on generating rank-reduced approximations of corresponding learned weight matrices of the model; and
update (308), the model to identify an origin of a synthetic content generated by the one or more new synthetic origins while maintaining detection capabilities of the model for the one or more synthetic origins.

12. A non-transitory computer-readable storage medium having stored thereon, a set of computer-executable instructions causing a computer comprising one or more processors to perform steps comprising:
receiving (302), a model, wherein the model is trained using training data associated with one or more synthetic origins;
receiving (304), a new training data associated with one or more new synthetic origins;
selectively training (306), one or more portions of the model utilizing the new training data, wherein the one or more portions of the model is selected based on generating rank-reduced approximations of corresponding learned weight matrices of the model; and
updating (308), the model to identify an origin of a synthetic content generated by the one or more new synthetic origins while maintaining detection capabilities of the model for the one or more synthetic origins.