Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
DETAILED ACTION
The United States Patent & Trademark Office appreciates the application that is submitted by the inventor/assignee. The United States Patent & Trademark Office reviewed the following application and has made the following comments below.
Information Disclosure Statement
The information disclosure statement (IDS) submitted on 05/21/2025 is considered and attached.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 3-5 and 10-11 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention. The examiner strongly suggested that appropriate corrections be made to clarify the claim scope.
With respect to Claim 3, the claim recites the following, each of which renders the claim indefinite:
“the synthetic media” on line 2 (unclear antecedent basis, the Examiner only finds “a synthetic media element” before this term. For the purpose the examination, the term is interpreted as “the synthetic media element”).
Claims 4-5 are also rejected due to their dependence on rejected independent claim 3.
With respect to Claim 10, the claim recites the following, each of which renders the claim indefinite:
“the deepfake” on line 2 (unclear antecedent basis, the Examiner only finds “a deepfake model” before this term. For the purpose the examination, the term is interpreted as “the deepfake model”).
With respect to Claim 11, the claim recites the following, each of which renders the claim indefinite:
“the media element comprises a video, wherein the video or image is divided” on line 2 (unclear antecedent basis for the “image”. For the purpose the examination, it is interpreted as “the media element comprises a video or a video, wherein the video or image is divided”).
Claim Rejections - 35 USC § 102
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claims 19, 21 and 24 are rejected under 35 U.S.C. 102 as being anticipated by Gandhi et al. (Gandhi, Kashish, et al. "A multimodal framework for deepfake detection." arXiv, published 10/04/2024, hereinafter Gandhi)
CLAIM 19
Regarding Claim 19, Gandhi teaches A computer (Gandhi, page 17, section 5: “pro cessors using i5 12th generation with Nvidia rtx 3050 16gbram”) implemented method of explainability of a media element identified as deepfake (Gandhi, page 5, section 4, see FIG. 1. Gandhi teaches a multimodal framework that can classify a video as deepfake or non-deepfake, the framework can explain which feature contributes for the classification result), comprising:
feeding the media element into each of a plurality of deepfake models of an ensemble (Gandhi, pages 13-14, section 4.2.1 and 4.2.2: “the videos were processed through our feature e traction model, resulting in a final feature dataframe with 2,590x13 … The data was subsequently split into training and testing sets in a 80:20 ratio. This was then fed to the various models mentioned below … Decision Trees … Random Forest … XGBoost was used as it focuses on patterns that are difficult to classify … Artifical Neural Network (ANN) for DeepFake video detection” Gandhi teaches feeding image data and audio data of input video to a group of machine learning and deep learning models to determine whether the data is fake), each deepfake model trained to analyze a specific feature for classifying the media element as deepfake or not deepfake (Gandhi, pages 13-14, section 4.2.1 and 4.2.2. Gandhi teaches specific models for image data and audio data. In addition, Gandhi teaches using different models for different facial features (such as skin tone, head pose, nose and lip, pupils,…) in section 4.1);
identifying a sub-set of the plurality of deepfake models of the ensemble that classified the media element as deepfake (Gandhi, page 5, section 4 Proposed methodology, see annotated FIG. 1 below. Gandhi teaches reporting which models output fake result, video models or audio models.);
PNG
media_image1.png
735
1050
media_image1.png
Greyscale
PNG
media_image2.png
742
1079
media_image2.png
Greyscale
and generating a presentation comprising specific features indicated with respect to the media element, wherein the specific features that are indicated correspond to the sub-set of the plurality of deepfake models that classified the media element as deepfake. (Gandhi, page 11, see annotated FIG. 7 below. Diagram of a SHAP analysis is plotted to how much each feature contributed to the classification result of the video models)
CLAIM 21
Regarding claim 21, Gandhi teaches the method of claim 19. In addition, Gandhi teaches the presentation includes markings and/or (The Examiner notes since a listing with “or” is disjunctive, any one of the elements found in the prior art is sufficient to reject the claim. While citations have been provided for completeness and rapid prosecution, only one element is required.) explanations for plurality of different specific features corresponding to a sub-set of the plurality of deepfake models of the ensemble classifying the media element as deepfake. (Gandhi, page 11, see annotated FIG. 7 above. Diagram of a SHAP analysis is plotted to show how much each feature contributed to the classification result of the video models)
CLAIM 24
Regarding claim 24, Gandhi teaches the method of claim 19. In addition, Gandhi teaches the presentation includes a text explanation of the specific features that the sub-set of the plurality of deepfake models that detected deepfake analyze. (Gandhi, page 11, see annotated FIG. 7 above: “The features include cheek bone height, inter-pupil distance, number of blinks, head pose angles (x, y, z), nose size, lip size, contrast correlation, luminance, chrominance1, chrominance2, and others, listed from 1 to 13, respectively”. Diagram of a SHAP analysis is plotted to show how much each feature contributed to the classification result of the video models)
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
CLAIM 20
Claim(s) 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Gandhi in view of Jia et al. (Jia, Xi, and Linlin Shen. "Skin lesion classification using class activation map." arXiv, published 2017, hereinafter Jia), and further in view of Price et al. (US-20220237799-A1, hereinafter Price_2022).
In regards to Claim 20, Gandhi teaches the method of Claim 19. In addition, Gandhi teaches detect faces in an image. (Gandhi, page 5, section 4.1.1: “the Haar Cas cade algorithm, since it can easily detect objects in images irrespective of their scale and location. Hence, it was employed to accurately identify faces within the frame, defining our region of interest (ROI).”)
Gandhi does not explicitly disclose Gandhi for each specific feature corresponding to a specific deepfake model: analyzing the media element to detect a plurality of instances of the specific feature; automatically dividing the media element into a plurality of sub-regions each include a respective instance of the specific feature; for each respective sub-region of the plurality of sub-regions, creating a synthetic media element including the respective sub-region and excluding other sub-regions;
Price_2022 is in the same field of art of classification system. Further, Price_2022 teaches for each specific feature corresponding to a specific deepfake model: analyzing the media element to detect a plurality of instances of the specific feature (Price_2022, ¶ [0085-0086] “the first object segmentation model 304 includes an object detection neural network … the object detection neural network 410 will accurately detect objects corresponding to known object classes … the object detection neural network 410 is a CNN (e.g., an R-CNN or a Faster R-CNN)” Price_2022 teaches detect objects using CNN); automatically (Price_2022, Abstract: “a multi-model object segmentation framework for automatically segmenting objects in digital images”) dividing the media element into a plurality of sub-regions each include a respective instance of the specific feature (Price_2022, ¶ [0069-0070]: “receiving an input image … the image includes one or more objects … the first object segmentation model 304 is an object segmentation machine-learning model trained to detect and classify objects corresponding to known object classes … the second object segmentation model 306 is an object segmentation machine-learning model trained to segment semantic objects and instances of objects in a panoptic manner (e.g., identifies all foreground and background objects)” Price_2022 teaches segmenting an images into different pixel region for different object); for each respective sub-region of the plurality of sub-regions, creating a synthetic media element including the respective sub-region and excluding other sub-regions (Price_2022, ¶ [0069-0070]: “the first object segmentation model 304 is a general object segmentation model that generates a first object mask set of objects in the input image … the second object segmentation model is a general object segmentation model that generates a second object mask set”, ¶ [0092]: “the object mask generator neural network 418 creates a separate image layer that sets the pixels corresponding to each detected object to positive (e.g., binary 1) while setting the remaining pixels in the image to a neutral or negative (e.g., binary 0). When an object mask layer for a target object is combined with the input image 302, only the pixels of the target object are visible. Indeed, the generated object mask facilitates the automatic selection of target objects within the input image 302” Price_2022 teaches generating object masks for classified objects, an object mask combined with input image will return pixels of corresponding object);
Therefore, it would have been obvious to one having ordinary skill in the art before the effective filing date of the claimed invention to modify the invention of Gandhi by incorporating object detection and segmentation that is taught by Price_2022, to make a face detection system that can perform detection and segmentation automatically; thus, one of ordinary skilled in the art would be motivated to combine the references since among its several aspects, the present invention recognizes there is a need to improve efficiency and consistency of the system. (Price_2022, ¶ [2022]: “the multi-model object segmentation system improves accuracy, flexibility, and efficiency utilizing a multi-model object segmentation framework that accurately and automatically segments objects in digital images”)
The combination of Gandhi and Price_2022 does not explicitly disclose does not explicitly disclose feeding an image into a classification model, then feeding a segmented object in the image into the classification model.
Jia is in the same field of art of classification model. Further, Jia teaches feeding an image into a classification model (Jia, page 1, section 2.1: “we input a skin lesion image into CNN and get its class activation maps in stage 1”, see annotated FIG. 1 below, stage 1), then feeding a segmented object in the image into the classification model. (Jia, page 1, section 2.1: “then we crop the import regions from original images according to CAM and use them as the input of stage 2 to produce the final probabilities”, see annotated FIG. 1 below, stage 2)
PNG
media_image3.png
529
1978
media_image3.png
Greyscale
PNG
media_image4.png
145
889
media_image4.png
Greyscale
Therefore, it would have been obvious to one having ordinary skill in the art before the effective filing date of the claimed invention to modify the invention of Ye by incorporating the two stages framework for classifying an image that is taught by Jia, to make an image classification system that attend both global and local features; thus, one of ordinary skilled in the art would be motivated to combine the references since among its several aspects, the present invention recognizes there is a need to improve performance of the classification model (Jia, Abstract: “The two stage framework achieved a mean AUC of 0.857 in ISIC-2017 skin lesion validation set and is 0.04 higher than that of the original inputs, 0.821.” The Examiner notes AUC is a measure for performance of classification, below is a definition of AUC from Google).
The combination of Gandhi, Price_2022 and Jia then teaches feeding each of a plurality of synthetic media elements (Price_2022, ¶ [0004]: “the disclosed systems return one or more of the segmented objects (or sub-objects)”) into the corresponding specific deepfake model (Gandhi, pages 13-14, section 4.2.1. Gandhi teaches various models to analyze facial features to determine a deepfake result) (Jia, page 1, section 2.1: “then we crop the import regions from original images according to CAM and use them as the input of stage 2 to produce the final probabilities”, see annotated FIG. 1 above, stage 2.),
PNG
media_image5.png
943
1123
media_image5.png
Greyscale
wherein the presenting indicates the instance of the specific feature corresponding to the sub-region of the synthetic media element identified by the deepfake model as likely deepfake. (Jia, page 2, section 2.3 Class Activation Map, see reconstructed text below. Jia teaches techniques to produce an explanation map (heatmap) indicates target object in an image)
Thus, the claimed subject matter would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention.
CLAIM 22
Claim(s) 22 is/are rejected under 35 U.S.C. 103 as being unpatentable over Gandhi in view of Adeel et al. (US-20240046515-A1, published 02/2024, hereinafter Adeel)
PNG
media_image6.png
734
1313
media_image6.png
Greyscale
Regarding claim 22, Gandhi teaches the method of Claim 19. In addition, Gandhi teaches the media element comprises a video (Gandhi, page 4, section 3.1 Video dataset), the specific feature includes a specific object depicted in the video (Gandhi, pages 5-6, section 4.1.1, see modified FIG. 2 below. Gandhi teaches detect face, then facial features are extracted and analyzed to classify the detected face).
Gandhi does not explicitly disclose the presentation includes the specific object marked on the video.
Adeel is in the same field of art of object classification system. Further, Adeel teaches the presentation includes the specific object marked on the video. (Adeel, ¶ [0020, 0035 and 0037]: “0020: The application of a machine-learning model may result in the content processing engine 104 drawing an outline shape that surrounds each object of the specific object type that is detected, as well as presenting a text label that indicates the object type of each object detected; 0037: Once an object of a specific object type is detected, the detection module 514 may superimpose an indicator on the video frame to show that the object of the specific object type is detected. For example, the indicator may include an outline shape that surrounds the image of the object, as well as present a text label that indicates the object type of the object”; see FIG. 2, outline 204. Adeel teaches detect and classify objects in video, and display a graphic indicator indicates the classified object and its class onto the video)
Therefore, it would have been obvious to one having ordinary skill in the art before the effective filing date of the claimed invention to modify the invention of Gandhi by incorporating the video content processing system that is taught by Adeel, to make an object classification system that can display graphic indicator of classified objects; thus, one of ordinary skilled in the art would be motivated to combine the references since among its several aspects, the present invention recognizes there is a need to provide visual explainability to a classification system and improve user experience. (Adeel, ¶ [0001]: “Given the large amount of such video content, it is often very tedious and time-consuming to identify the exact video content that is being sought”)
Thus, the claimed subject matter would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention.
CLAIM 23
Claim(s) 23 is/are rejected under 35 U.S.C. 103 as being unpatentable over Gandhi in view of Govindu et al. (Govindu, A., et al. "Deepfake audio detection and justification with Explainable Artificial Intelligence (XAI).", published 2023, hereinafter Govindu).
In regards to Claim 23, Gandhi teaches the method of Claim 19. In addition, Gandhi teaches the media element comprises audio (Gandhi, page 4, section 3.2 Audio Dataset: “For audio dataset, the Fake-or-Real (FoR) dataset [26] was utilized, which com prises over 195,000 utterances from both real human speech and computer generated synthetic speech …”), the specific feature includes a specific audio feature (Gandhi, pages 11-12, section 4.1.3 For Audio: “To understand and visually study audio signals, mel-spectrograms were utilized. “Mel” is an abbreviation for “melody.” Mel-spectrograms [27] provide a time-frequency representation of audio signals with perceptually relevant amplitude and frequency representations”).
Gandhi does not explicitly disclose the presentation includes an explanation of the specific audio feature or playing the specific audio feature over speakers.
Govindu is in the same field of art of detecting deepfake audio. Further, Govindu teaches the presentation includes an explanation of the specific audio feature (Govindu, pages 13-14, section 5.3: “We used three different XAI techniques, LIME, GradCAM and SHAP, to understand how each model was making its predictions. SpecificaJJy, we used these techniques to identify frequency bands or pixels of the spectrogram that were critical for each prediction, and to compare the importance of different features across different models... We tested all models on a Spectrogram and viewed SHAPiey values for the image, as illustrated in column l of Table 6. Green regions indicate maximum positive impact pixels on classification, while red regions denote maximum negative impact” See FIG. 8 and Table 6. Govindu teaches generating fake audio samples, creating the mel-spectrogram for said samples, using deep learning models to classify the samples as fake or real and explaining the results by LIME, GRADcam or SHAP) or (The Examiner notes since a listing with “or” is disjunctive, any one of the elements found in the prior art is sufficient to reject the claim. While citations have been provided for completeness and rapid prosecution, only one element is required.) playing the specific audio feature over speakers.
Therefore, it would have been obvious to one having ordinary skill in the art before the effective filing date of the claimed invention to modify the invention of Gandhi by incorporating method to explain the audio deepfake detection that is taught by Govindu, to make a system not only detect deepfake audio but also provide explainability for the detection; thus, one of ordinary skilled in the art would be motivated to combine the references since among its several aspects, the present invention recognizes there is a need to provide explainability to deep learning models (Govindu, page 2, third paragraph: “We use XAI that comprises tools and techniques to detect, study and explain the "thought process" of a computational model. It converts a deep learning model to a white box model, by highlighting parameters the model learnt and interpreting model predictions (Guo, 2020). It focuses on explanation, meaningfulness, accuracy, and knowledge limits of the model”).
Thus, the claimed subject matter would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention.
Claim(s) 1, 2, 6, 7, 9, 12, 15, 17 and 25 is/are rejected under 35 U.S.C. 103 as being unpatentable over Ye (WO-2021080815-A1, a copy of this document is attached, hereinafter Ye) in view of Jia et al. (Jia, Xi, and Linlin Shen. "Skin lesion classification using class activation map." arXiv, published 2017, hereinafter Jia), and further in view of Hua et al. (CN-118261255-A, published 06/2024, a translated document is attached, hereinafter Hua).
CLAIM 1
In regards to Claim 1, Ye teaches a computer implemented method (Ye, ¶ [0035]: “computer-implemented steps for processing information in the system. Instructions can be implemented in software, firmware or hardware and include any type of programmed step undertaken by components of the system”) of explainability of a media element identified as deepfake (Ye, ¶ [0005]: “Accordingly, techniques are provided herein to determine whether a video is genuine or is a fake generated by machine learning”), comprising:
feeding the media element into a deepfake model trained to classify the media element as deepfake or not deepfake (Ye, ¶ [0062-0068]: “at block 300, an image is received. The image can be directly analyzed at block 302 by processing the image through a first neural network (NN) such as a convolutional NN (CNN) … an indication may be returned at block 316 that the image is fake … The fake images may be generated using "deepfake" techniques from the ground truth original images by the designer”, see modified FIG. 3 below),
wherein the deepfake model is implemented as a non-explainable model (Ye, ¶ [0062, 0065 and 0073-0075]: “The image can be directly analyzed at block 302 by processing the image through a first neural network (NN) such as a convolutional NN (CNN)” The Examiner notes CNN is a black box model, Ye does not disclose using any explainable techniques such as saliency map or class activation map);
PNG
media_image7.png
598
800
media_image7.png
Greyscale
Ye does not explicitly disclose feeding an image into a classification model, then feeding a segmented object in the image into the classification model.
Jia is in the same field of art of classification model. Further, Jia teaches feeding an image into a classification model (Jia, page 1, section 2.1: “we input a skin lesion image into CNN and get its class activation maps in stage 1”, see annotated FIG. 1 below, stage 1), then feeding a segmented object in the image into the classification model. (Jia, page 1, section 2.1: “then we crop the import regions from original images according to CAM and use them as the input of stage 2 to produce the final probabilities”, see annotated FIG. 1 below, stage 2)
PNG
media_image3.png
529
1978
media_image3.png
Greyscale
PNG
media_image4.png
145
889
media_image4.png
Greyscale
Therefore, it would have been obvious to one having ordinary skill in the art before the effective filing date of the claimed invention to modify the invention of Ye by incorporating the two stages framework for classifying an image that is taught by Jia, to make an image classification system that attend both global and local features; thus, one of ordinary skilled in the art would be motivated to combine the references since among its several aspects, the present invention recognizes there is a need to improve performance of the classification model (Jia, Abstract: “The two stage framework achieved a mean AUC of 0.857 in ISIC-2017 skin lesion validation set and is 0.04 higher than that of the original inputs, 0.821.” The Examiner notes AUC is a measure for performance of classification, below is a definition of AUC from Google).
The combination of Ye and Hua does not explicitly disclose in response to the classification model classifying the media element: automatically dividing the media element into a plurality of sub-regions; for each respective sub-region of the plurality of sub-regions, creating a synthetic media element including the respective sub-region and excluding other sub-regions;
Hua is in the same field of art of classification model. Further, Hua teaches in response to the classification model classifying the media element (Hua, ¶ [n0055-n0056]: “Step 1: Input image I and its corresponding prior saliency map … The deep models are primarily explained using a CNN-based architecture… The methods for calculating the prior saliency map include, but are not limited to, white-box based methods: Saliency, Grad-CAM, Grad-CAM++, ScoreCAM, and black-box based methods: LIME, Kernel Shap, RISE, and HSIC-Attribution.” An image is classified by a CNN-based model): automatically dividing the media element into a plurality of sub-regions (Hua, ¶ [n0042]: “The image segmentation module is used to divide image I into m sub-regions using a prior saliency map and construct an element set.”; ¶ [n0057]: “Using sub-region segmentation, in order to obtain interpretable regions in the image, the image is divided into M sub-regions IM”); for each respective sub-region of the plurality of sub-regions, creating a synthetic media element including the respective sub-region and excluding other sub-regions (Hua, ¶ [n0057]: “Using sub-region segmentation, in order to obtain interpretable regions in the image, the image is divided into M sub-regions IM, where M represents the sub-region IM formed by covering a portion of image I”. Hua teaches segmenting the input image I into multiple regions, and create a plurality of images IM by masking corresponding segmented regions);
Therefore, it would have been obvious to one having ordinary skill in the art before the effective filing date of the claimed invention to modify the invention of Ye and Jia by incorporating image segmentation module that is taught by Hua, to make a system that can segment image automatically; thus, one of ordinary skilled in the art would be motivated to combine the references since among its several aspects, the present invention recognizes there is a need to substitute Jia’s manual segmenting and cropping with automatic segmentation method, such the change from manual to automatic will improve efficiency and consistency of the system.
The combination of Ye, Jia and Hua teaches in response to the deepfake model classifying the media element (Hua, ¶ [n0055-n0056]: “Step 1: Input image I and its corresponding prior saliency map … The deep models are primarily explained using a CNN-based architecture… The methods for calculating the prior saliency map include, but are not limited to, white-box based methods: Saliency, Grad-CAM, Grad-CAM++, ScoreCAM, and black-box based methods: LIME, Kernel Shap, RISE, and HSIC-Attribution.” An image is classified by a CNN-based model) as deepfake (Ye, ¶ [0062-0068]: “at block 300, an image is received. The image can be directly analyzed at block 302 by processing the image through a first neural network (NN) such as a convolutional NN (CNN) … an indication may be returned at block 316 that the image is fake … The fake images may be generated using "deepfake" techniques from the ground truth original images by the designer”, see modified FIG. 3 above): automatically dividing the media element into a plurality of sub-regions (Hua, ¶ [n0042]: “The image segmentation module is used to divide image I into m sub-regions using a prior saliency map and construct an element set.”; ¶ [n0057]: “Using sub-region segmentation, in order to obtain interpretable regions in the image, the image is divided into M sub-regions IM”); for each respective sub-region of the plurality of sub-regions, creating a synthetic media element including the respective sub-region and excluding other sub-regions (Hua, ¶ [n0057]: “Using sub-region segmentation, in order to obtain interpretable regions in the image, the image is divided into M sub-regions IM, where M represents the sub-region IM formed by covering a portion of image I”. Hua teaches segmenting the input image I into multiple regions, and create a plurality of images IM by masking corresponding segmented regions); feeding each of a plurality of synthetic media elements into the deepfake model (Jia, page 1, section 2.1: “then we crop the import regions from original images according to CAM and use them as the input of stage 2 to produce the final probabilities”, see annotated FIG. 1 above, stage 2); in response to the deepfake model classifying at least one of the synthetic media elements as being deepfake (Ye, ¶ [0063-0067]: “the image may be input to a face recognition module to analyze for artifacts, also referred to herein as irregularities, in the face and/or background of the image, as well as lighting irregularities in the image … an irregularity in a face in the image (spatial domain) may include a small region having a checkerboard-like appearance, indicating blurry resolution owing to digital altering.” Ye teaches … if any irregularity exists in any domain, an indication may be returned at block 316 that the image is fake” Ye teaches determining an image is fake in response to detecting an irregular region in the image), creating a presentation indicating the sub-region corresponding to each at least one synthetic media elements classified as deepfake (Hua, ¶ [n0056]: “The methods for calculating the prior saliency map include, but are not limited to, white-box based methods: Saliency, Grad-CAM, Grad-CAM++, ScoreCAM, and black-box based methods: LIME, Kernel Shap, RISE, and HSIC-Attribution”; Jia, page 2, section 2.3 Class Activation Map, see reconstructed text below. The Examiner notes both Hua and Jia teaches techniques to produce an explanation map (heatmap) indicates target object in an image).
PNG
media_image5.png
943
1123
media_image5.png
Greyscale
Thus, the claimed subject matter would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention.
CLAIM 2
Regarding Claim 2, the combination of Ye, Jia and Hua teaches the method of Claim 1. In addition, the combination of Ye, Jia and Hua teaches the synthetic media element includes the respective sub-region of the media element (Hua, ¶ [n0057]: “Using sub-region segmentation, in order to obtain interpretable regions in the image, the image is divided into M sub-regions IM, where M represents the sub-region IM formed by covering a portion of image I”), and values denoting lack of information for the excluded sub-regions. (Hua, ¶ [n0058]: “an existing image attribution algorithm is applied to compute the saliency map for the corresponding category I. After that, the map is adjusted to an N×N size, where the values represent the importance of each patch…Based on the determined importance of each patch, d patches are sequentially assigned to each sub-region IM, while the remaining patch regions are covered by 0” Hua teaches masking the pixel intensity of excluded regions as 0)
CLAIM 6
Regarding Claim 6, the combination of Ye, Jia and Hua teaches the method of Claim 1. In addition, the combination of Ye, Jia and Hua teaches the deepfake model generates a classification in response to an input and excludes an indication of specific features of the input that most significantly contributed to the classification. (Ye, ¶ [0062-0068]: “at block 300, an image is received. The image can be directly analyzed at block 302 by processing the image through a first neural network (NN) such as a convolutional NN (CNN) … an indication may be returned at block 316 that the image is fake … The fake images may be generated using "deepfake" techniques from the ground truth original images by the designer”, see modified FIG. 3 above. Ye teaches using CNN-based model to classify the image/video as fake or non-fake, not including generating any map or presentation indicates which features contributed to the classification)
CLAIM 7
Regarding Claim 7, the combination of Ye, Jia and Hua teaches the method of Claim 1. In addition, the combination of Ye, Jia and Hua teaches the synthetic media element is generated in a format designed to comply with requirements for input into the deepfake model. (Hua, ¶ [n0057]: “Using sub-region segmentation, in order to obtain interpretable regions in the image, the image is divided into M sub-regions IM, where M represents the sub-region IM formed by covering a portion of image I”. Hua generating plurality images IM which also are image-format as input image) (Jia, page 1, section 2.1: “we input a skin lesion image into CNN and get its class activation maps in stage 1. And then we crop the import regions from original images according to CAM and use them as the input of stage 2 to produce the final probabilities”, see modified FIG. 1 below. Jia teaches the segmented and cropped image is input to the same classification model as the original input image)
PNG
media_image8.png
530
1095
media_image8.png
Greyscale
CLAIM 9
PNG
media_image8.png
530
1095
media_image8.png
Greyscale
Regarding Claim 9, the combination of Ye, Jia and Hua teaches the method of Claim 1. In addition, the combination of Ye, Jia and Hua teaches the synthetic media element is created for avoiding or reducing likelihood of the deepfake model classifying the synthetic media element as deepfake based on the excluded sub-regions. (Jia, page 1, section 2.1: “we input a skin lesion image into CNN and get its class activation maps in stage 1. And then we crop the import regions from original images according to CAM and use them as the input of stage 2 to produce the final probabilities”, see modified FIG. 1 below. Jia teaches segmenting the lesion, removing the background region around the lesion, so the classification network can focus on the lesion region and ignoring the removed background region)
CLAIM 12
Regarding Claim 12, the combination of Ye, Jia and Hua teaches the method of Claim 1. In addition, the combination of Ye, Jia and Hua teaches the media element comprises a video or (***The Examiner notes since a listing with “or” is disjunctive, any one of the elements found in the prior art is sufficient to reject the claim. While citations have been provided for completeness and rapid prosecution, only one element is required.) image (Ye, ¶ [0060]: “a first technique for determining whether an image is "fake", i.e., has been digitally altered from an original image, is illustrated … The image 200 may be an image such as an I-frame from a video stream, and some or all of the frames of the video stream may be processed as disclosed herein.”), wherein the video or image is divided into a plurality of sub-regions each including an area of a frame of the video or of the image. (Hua, ¶ [n0042]: “The image segmentation module is used to divide image I into m sub-regions using a prior saliency map and construct an element set.”; ¶ [n0057]: “Using sub-region segmentation, in order to obtain interpretable regions in the image, the image is divided into M sub-regions IM”. Hua teaches segmenting an image into multiple sub-region images, each includes an area of the original image)
CLAIM 15
Regarding Claim 15, the combination of Ye, Jia and Hua teaches the method of Claim 1. In addition, the combination of Ye, Jia and Hua teaches the presentation includes a marking on the media element (Hua, ¶ [n0056]: “The methods for calculating the prior saliency map include, but are not limited to, white-box based methods: Saliency, Grad-CAM, Grad-CAM++, ScoreCAM, and black-box based methods: LIME, Kernel Shap, RISE, and HSIC-Attribution”; Jia, page 2, section 2.3 Class Activation Map, see reconstructed text below. The Examiner notes both Hua and Jia teaches generating a heat map for classification result)., each sub-region classified as deepfake. (Ye, ¶ [0062-0068]: “at block 300, an image is received. The image can be directly analyzed at block 302 by processing the image through a first neural network (NN) such as a convolutional NN (CNN) … an indication may be returned at block 316 that the image is fake … The fake images may be generated using "deepfake" techniques from the ground truth original images by the designer”)
PNG
media_image5.png
943
1123
media_image5.png
Greyscale
CLAIM 17
Regarding Claim 17, the combination of Ye, Jia and Hua teaches the method of Claim 1. In addition, the combination of Ye, Jia and Hua teaches processing the media element and/or processing each sub-region by at least one of: applying a filter (Hua, ¶ [n0057]: “Using sub-region segmentation, in order to obtain interpretable regions in the image, the image is divided into M sub-regions IM, where M represents the sub-region IM formed by covering a portion of image I”, ¶ [n0058]: “an existing image attribution algorithm is applied to compute the saliency map for the corresponding category I. After that, the map is adjusted to an N×N size, where the values represent the importance of each patch…Based on the determined importance of each patch, d patches are sequentially assigned to each sub-region IM, while the remaining patch regions are covered by 0” Hua teaches performing segmentation by applying a mask/filter to the image ) and/or (The Examiner notes since a listing with “or” is disjunctive, any one of the elements found in the prior art is sufficient to reject the claim. While citations have been provided for completeness and rapid prosecution, only one element is required.) changing from color to greyscale, wherein the plurality of synthetic media elements are created based on the processed sub-regions.
CLAIM 25
In regards to Claim 1, Ye teaches A system for explainability of a media element identified as deepfake (Ye, ¶ [0005-0006]: “techniques are provided herein to determine whether a video is genuine or is a fake generated by machine learning … ”), comprising: at least one processor (Ye, ¶ [0013]: “an apparatus includes at least one computer storage medium with instructions executable by at least one processor”) executing a code for:
feeding the media element into a deepfake model trained to classify the media element as deepfake or not deepfake (Ye, ¶ [0062-0068]: “at block 300, an image is received. The image can be directly analyzed at block 302 by processing the image through a first neural network (NN) such as a convolutional NN (CNN) … an indication may be returned at block 316 that the image is fake … The fake images may be generated using "deepfake" techniques from the ground truth original images by the designer”, see modified FIG. 3 below),
wherein the deepfake model is implemented as a non-explainable model (Ye, ¶ [0062, 0065 and 0073-0075]: “The image can be directly analyzed at block 302 by processing the image through a first neural network (NN) such as a convolutional NN (CNN)” The Examiner notes CNN is a black box model, Ye does not disclose using any explainable techniques such as saliency map or class activation map);
PNG
media_image7.png
598
800
media_image7.png
Greyscale
Ye does not explicitly disclose feeding an image into a classification model, then feeding a segmented object in the image into the classification model.
Jia is in the same field of art of classification model. Further, Jia teaches feeding an image into a classification model (Jia, page 1, section 2.1: “we input a skin lesion image into CNN and get its class activation maps in stage 1”, see annotated FIG. 1 below, stage 1), then feeding a segmented object in the image into the classification model. (Jia, page 1, section 2.1: “then we crop the import regions from original images according to CAM and use them as the input of stage 2 to produce the final probabilities”, see annotated FIG. 1 below, stage 2)
PNG
media_image3.png
529
1978
media_image3.png
Greyscale
Therefore, it would have been obvious to one having ordinary skill in the art before the effective filing date of the claimed invention to modify the invention of Ye by incorporating the two stages framework for classifying an image that is taught by Jia, to make an image classification system that attend both global and local features; thus, one of ordinary skilled in the art would be motivated to combine the references since among its several aspects, the present invention recognizes there is a need to improve performance of the classification model (Jia, Abstract: “The two stage framework achieved a mean AUC of 0.857 in ISIC-2017 skin lesion validation set and is 0.04 higher than that of the original inputs, 0.821.” The Examiner notes AUC is a measure for performance of classification, below is a definition of AUC from Google).
PNG
media_image4.png
145
889
media_image4.png
Greyscale
The combination of Ye and Hua does not explicitly disclose in response to the classification model classifying the media element: automatically dividing the media element into a plurality of sub-regions; for each respective sub-region of the plurality of sub-regions, creating a synthetic media element including the respective sub-region and excluding other sub-regions;
Hua is in the same field of art of classification model. Further, Hua teaches in response to the classification model classifying the media element (Hua, ¶ [n0055-n0056]: “Step 1: Input image I and its corresponding prior saliency map … The deep models are primarily explained using a CNN-based architecture… The methods for calculating the prior saliency map include, but are not limited to, white-box based methods: Saliency, Grad-CAM, Grad-CAM++, ScoreCAM, and black-box based methods: LIME, Kernel Shap, RISE, and HSIC-Attribution.” An image is classified by a CNN-based model): automatically dividing the media element into a plurality of sub-regions (Hua, ¶ [n0042]: “The image segmentation module is used to divide image I into m sub-regions using a prior saliency map and construct an element set.”; ¶ [n0057]: “Using sub-region segmentation, in order to obtain interpretable regions in the image, the image is divided into M sub-regions IM”); for each respective sub-region of the plurality of sub-regions, creating a synthetic media element including the respective sub-region and excluding other sub-regions (Hua, ¶ [n0057]: “Using sub-region segmentation, in order to obtain interpretable regions in the image, the image is divided into M sub-regions IM, where M represents the sub-region IM formed by covering a portion of image I”. Hua teaches segmenting the input image I into multiple regions, and create a plurality of images IM by masking corresponding segmented regions);
Therefore, it would have been obvious to one having ordinary skill in the art before the effective filing date of the claimed invention to modify the invention of Ye and Jia by incorporating image segmentation module that is taught by Hua, to make a system that can segment image automatically; thus, one of ordinary skilled in the art would be motivated to combine the references since among its several aspects, the present invention recognizes there is a need to substitute Jia’s manual segmenting and cropping with automatic segmentation method, such the change from manual to automatic will improve efficiency and consistency of the system.
The combination of Ye, Jia and Hua teaches in response to the deepfake model classifying the media element (Hua, ¶ [n0055-n0056]: “Step 1: Input image I and its corresponding prior saliency map … The deep models are primarily explained using a CNN-based architecture… The methods for calculating the prior saliency map include, but are not limited to, white-box based methods: Saliency, Grad-CAM, Grad-CAM++, ScoreCAM, and black-box based methods: LIME, Kernel Shap, RISE, and HSIC-Attribution.” An image is classified by a CNN-based model) as deepfake (Ye, ¶ [0062-0068]: “at block 300, an image is received. The image can be directly analyzed at block 302 by processing the image through a first neural network (NN) such as a convolutional NN (CNN) … an indication may be returned at block 316 that the image is fake … The fake images may be generated using "deepfake" techniques from the ground truth original images by the designer”, see modified FIG. 3 above): automatically dividing the media element into a plurality of sub-regions (Hua, ¶ [n0042]: “The image segmentation module is used to divide image I into m sub-regions using a prior saliency map and construct an element set.”; ¶ [n0057]: “Using sub-region segmentation, in order to obtain interpretable regions in the image, the image is divided into M sub-regions IM”); for each respective sub-region of the plurality of sub-regions, creating a synthetic media element including the respective sub-region and excluding other sub-regions (Hua, ¶ [n0057]: “Using sub-region segmentation, in order to obtain interpretable regions in the image, the image is divided into M sub-regions IM, where M represents the sub-region IM formed by covering a portion of image I”. Hua teaches segmenting the input image I into multiple regions, and create a plurality of images IM by masking corresponding segmented regions); feeding each of a plurality of synthetic media elements into the deepfake model (Jia, page 1, section 2.1: “then we crop the import regions from original images according to CAM and use them as the input of stage 2 to produce the final probabilities”, see annotated FIG. 1 above, stage 2); in response to the deepfake model classifying at least one of the synthetic media elements as being deepfake (Ye, ¶ [0063-0067]: “the image may be input to a face recognition module to analyze for artifacts, also referred to herein as irregularities, in the face and/or background of the image, as well as lighting irregularities in the image … an irregularity in a face in the image (spatial domain) may include a small region having a checkerboard-like appearance, indicating blurry resolution owing to digital altering.” Ye teaches … if any irregularity exists in any domain, an indication may be returned at block 316 that the image is fake” Ye teaches determining an image is fake in response to detecting an irregular region in the image), creating a presentation indicating the sub-region corresponding to each at least one synthetic media elements classified as deepfake (Hua, ¶ [n0056]: “The methods for calculating the prior saliency map include, but are not limited to, white-box based methods: Saliency, Grad-CAM, Grad-CAM++, ScoreCAM, and black-box based methods: LIME, Kernel Shap, RISE, and HSIC-Attribution”; Jia, page 2, section 2.3 Class Activation Map, see reconstructed text below. The Examiner notes both Hua and Jia teaches techniques to produce an explanation map (heatmap) indicates target object in an image).
PNG
media_image5.png
943
1123
media_image5.png
Greyscale
Thus, the claimed subject matter would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention.
Claim(s) 8 and 10 is/are rejected under 35 U.S.C. 103 as being unpatentable over Ye in view of Jia in view of Hua, and further in view of Price et al. (US-20220237799-A1, hereinafter Price_2022).
CLAIM 8
In regards to Claim 8, the combination of Ye, Jia and Hua teaches the method of Claim 1. In addition, Hua teaches segmentation using masking technique. (Hua, ¶ [n0057]: “Using sub-region segmentation, in order to obtain interpretable regions in the image, the image is divided into M sub-regions IM, where M represents the sub-region IM formed by covering a portion of image I”. Hua teaches segmenting the input image I into multiple regions, and create a plurality of images IM by masking corresponding segmented regions)
The combination of Ye, Jia and Hua does not explicitly disclose the synthetic media element is created by creating a mask corresponding to the respective sub-region, and applying the mask to the media element.
Price_2022 is in the same field of art of image segmentation. Further, Price_2022 teaches the synthetic media element is created by creating a mask corresponding to the respective sub-region, and applying the mask to the media element. (Price_2022, ¶ [0092-0093]: “In generating an object mask for a detected object, the object mask generator neural network 418 is able to segment the pixels in the detected query object from the other pixels in the image. For example, the object mask generator neural network 418 creates a separate image layer that sets the pixels corresponding to each detected object to positive (e.g., binary 1) while setting the remaining pixels in the image to a neutral or negative (e.g., binary 0)” Price teaches segmenting objects by create a mask/layer that set pixels of detected objects to 1, and set background pixels to 0.)
Therefore, it would have been obvious to one having ordinary skill in the art before the effective filing date of the claimed invention to modify the invention of Ye, Jia and Hua by incorporating the segmentation framework that is taught by Price_2022, to make a more robust object segmentation system; thus, one of ordinary skilled in the art would be motivated to combine the references since among its several aspects, the present invention recognizes there is a need to improve accuracy, flexibility, and efficiency (Price_2022, ¶ [0036-0039]: “the multi-model object segmentation system improves accuracy, flexibility, and efficiency utilizing a multi-model object segmentation framework that accurately and automatically segments objects in digital images”).
Thus, the claimed subject matter would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention.
CLAIM 10
In regards to Claim 10, the combination of Ye, Jia and Hua teaches the method of Claim 1.
The combination of Ye, Jia and Hua does not explicitly disclose the excluded sub-regions are generated for being non-significant for classification.
Price_2022 is in the same field of art of image segmentation. Further, Price_2022 teaches the excluded sub-regions are generated for being non-significant for classification. (Price_2022, ¶ [0070]: “ the second object segmentation model 306 is an object segmentation machine-learning model trained to segment semantic objects and instances of objects in a panoptic manner (e.g., identifies all foreground and background objects)”; ¶ [0152]: “the multi-model object segmentation system 106 performs the act 810 of generating a new unclassified object mask of the pixel cluster. For example, in one or more implementations, the multi-model object segmentation system 106 segments the pixel cluster into a new unclassified object mask. In additional implementations, the multi-model object segmentation system 106 classifies the object mask with a label of “unknown,” “unclassified,” “background,” “null,” or “no label””. Price_2022 teaches segmenting an images into multiple objects, including both foreground and background objects, the Examiner interprets “being non-significant for classification” as background/un-classified regions)
Therefore, it would have been obvious to one having ordinary skill in the art before the effective filing date of the claimed invention to modify the invention of Ye, Jia and Hua by incorporating the segmentation framework that is taught by Price_2022, to make a more robust object segmentation system; thus, one of ordinary skilled in the art would be motivated to combine the references since among its several aspects, the present invention recognizes there is a need to improve accuracy, flexibility, and efficiency (Price_2022, ¶ [0036-0039]: “the multi-model object segmentation system improves accuracy, flexibility, and efficiency utilizing a multi-model object segmentation framework that accurately and automatically segments objects in digital images”).
the excluded sub-regions are generated for being non-significant for classification (Price_2022, ¶ [0070]: “ the second object segmentation model 306 is an object segmentation machine-learning model trained to segment semantic objects and instances of objects in a panoptic manner (e.g., identifies all foreground and background objects)”; ¶ [0152]: “the multi-model object segmentation system 106 performs the act 810 of generating a new unclassified object mask of the pixel cluster. For example, in one or more implementations, the multi-model object segmentation system 106 segments the pixel cluster into a new unclassified object mask. In additional implementations, the multi-model object segmentation system 106 classifies the object mask with a label of “unknown,” “unclassified,” “background,” “null,” or “no label””. Price_2022 teaches segmenting an images into multiple objects, including both foreground and background objects, the Examiner interprets “being non-significant for classification” as background/un-classified regions) by the deepfake model. (Ye, ¶ [0062-0068]: “at block 300, an image is received. The image can be directly analyzed at block 302 by processing the image through a first neural network (NN) such as a convolutional NN (CNN) … an indication may be returned at block 316 that the image is fake … The fake images may be generated using "deepfake" techniques from the ground truth original images by the designer”)
Thus, the claimed subject matter would have been obvious to a person having ordinary skill in the art before the effective filing date of the claimed invention.
Allowable Subject Matter
Claims 3-5, 11, 13-14, 16 and 18 are objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
Pertinent Arts
The prior art made of record and not relied upon is considered pertinent to applicant’s disclosure.
Matthews et al. (US-20250039498-A1, filed 10/11/2024) which is directed to an apparatus includes interface circuitry to receive a media file, machine readable instructions, and at least one processor circuit to be programmed by the instructions to generate, based on a deepfake classification model, a classification score for the media file, obtain a class output from a final convolution feature map of the deepfake classification model, generate an explainability map based on a pooled weighted feature map along a channel dimension of the feature map, and identify the media file as a real or a deepfake media file based on the classification score and the explainability map.
Lee et al. (US-20260073732-A1, priority claimed 9/9/2024) which is directed to detect, in real-time, if a captured image and/or audio segment comprises a deepfake.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to NHUT HUY (JEREMY) PHAM whose telephone number is (703)756-5797. The examiner can normally be reached Mo - Fr. 8:30am - 6pm ET.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, O'Neal Mistry can be reached on (313)446-4912. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/NHUT HUY PHAM/Examiner, Art Unit 2674
/ONEAL R MISTRY/Supervisory Patent Examiner, Art Unit 2674
Prosecution Insights