DETAILED ACTION
This communication responsive to the Application No. 18/311,986 filed on 12/29/2025. Claims 1-20 are pending and are directed towards ADVERSE OR MALICIOUS INPUT MITIGATION FOR LARGE LANGUAGE MODELS
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 .
Response to Arguments
A new ground of rejection has been made based on the amendments. Ma discloses the core chatbot/generative-response flow of receiving natural-language user input through a chat UI, processing it with a VSDG/VAE-GAN dialog generator and displaying the generated response. Yang discloses a dynamic support selection where comparing current input embeddings to support embeddings and selecting closest examples using distance or cosine similarity. Bennett discloses stored question answer/example pairs and context-based subset selection where only a limited subset of Q/A pairs is made appropriate based on the user’s current context or section. Wysopal discloses updating security-response knowledge for new threats such as new attack vectors, threat metadata, remediation information, scans, workflows and reports are added and updated as threats evolve.
Combinedly, the combination discloses a generative chatbot that receives NL input, selects a context-relevant subset of stored example Q/A pairs using similarity/context filtering, uses those examples to guide response generation and updates the security and mitigation knowledge base as threats and attack vectors arise.
Claim Rejections - 35 USC § 103
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 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.
Claims 1, 2,3,6,7,11, 13,15,16,17, are rejected under 35 U.S.C. 103 as being unpatentable over Ma et al. (US 20200226475 A1), hereinafter referred to as Ma, in view of Yang et al. (US 20210374347 A1), hereinafter referred to as Yang.
As per claim 1, Ma discloses a system for implementing malicious input mitigation for artificial intelligence ("Al") and/or machine learning ("ML") models, the system comprising: a computing system, comprising:
at least one processor; and (A processor, Ma, para [0090])
a computer storage medium communicatively coupled to the at least one processor, the computer storage medium storing instructions that, when executed by the at least one processor, causes the computing system to perform operations comprising: (Store in non-transitory memory, Ma, para [0090])
receiving, via a user interface, a natural language ("NL") input during a communication session for generating responses to NL inputs by a generative Al system between the user interface and the computing system; (A system including a user interface, such as a computer, smart phone, tablet, etc., may allow a user to communicate with an artificial intelligence (AI) dialog generator, or chatbot, by entering query texts, Ma, para [0018]. A text box 202 of a chatbot may be presented where a query text may be entered by a user by typing on a keyboard or speaking into a microphone if the device is adapted with voice recognition capabilities and if query may be entered using natural language, Ma, para [0045]. One task of the NLP system is to generate natural responses to input queries using Natural Language Generation (NLG), herein referred to as dialog generation (DG), Ma, para [0002]. Here, the user interface is analogous to the computer and chat text box. The NL input aligns with the query text entered using natural language. The communication sessions are the user communicating with the AI dialog generator)
providing the prompt to the generative Al system, the prompt causing the generative Al system to generate an output that follows one of the example outputs in the subset of the example pairs; (Receiving an input query by a variational autoencoder (VAE), embedding the input query into vector representations via the VAE, converting the vector representations into responses by a generative adversarial network (GAN). Generating, with the GAN-G, a response sentence to the input query, Ma, para [0089]. Here, receiving the input query at the VAE and converting it into vector representations for response generation is analogous to providing the prompt)
receiving the output; (Receiving, at a generator of the GAN (GAN-G), the response embedding from the decoder, and generating, with the GAN-G, a response sentence to the input query, Ma, para [0089]. Here, the received output is analogous to the generated response sentence produced by the GAN-G and then displayed by the chatbot system)
generating a response to the NL input based on the received output; and (The VAE may convert queries received by the chatbot to vectors that are then used by the neural network to generate a response, the GAN-G generating a response to the received response embeddings, Ma, para [0018]. The response to the NL input is analogous to chatbot/ GAN generated response sentence generated from vectors response embeddings derived from the user query)
causing, via the user interface, presentation of the generated response within the communication session (Displaying a response to the input query in a user interface. The response is displayed in a response dialog box 206 below the query text box 202. The answer to the question is displayed on a computer screen by the chatbot for the user to observe. Ma, para [0046], [0083], [0089]).
However, Ma does not explicitly disclose the limitations:
dynamically identifying a subset of example pairs of example dialogue contexts containing malicious inputs and example outputs containing malicious input mitigation responses whose inputs are similar to inputs of a current dialogue context, the current dialogue context including the NL input;
generating a prompt, for [[a]] the generative Al system, comprising the current dialogue context with the NL input and the dynamically identified subset of example pairs;
Yang discloses:
dynamically identifying a subset of example pairs of example dialogue contexts containing malicious inputs and example outputs containing malicious input mitigation responses whose inputs are similar to inputs of a current dialogue context, the current dialogue context including the NL input; (Support text that provides a few examples compute an embedding vector for each word of the support text, compute an embedding vector for each word of input text and then computes distances between the embedding vectors. Where an input token and a support token have similar meanings or other similar aspects, the corresponding token embeddings are more likely to be close to each other, for each input token, determine which support token is closest to the input token and may compute a Euclidean distance or a cosine similarity, Yang, para [0094])
generating a prompt, for [[a]] the generative Al system, comprising the current dialogue context with the NL input and the dynamically identified subset of example pairs; (The inputs to system 500 includes support text and input text, support token are obtained and the support token may be multiple sequences of support tokens corresponding to text, such as multiple sentences, Yang, para [0046]. Here, the prompt is interpreted as a model input package. The current dialogue context with the NL input is analogous to current input text/query text. The dynamically identified subset of example pairs aligns with the support text and examples selected by similarity).
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma and Yang by updating of response generation by an AI chatbot (Ma) with named entity recognition (Yang). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma and Yang in order to efficiently search and access user-specific data items (See Yang, para [0046])
As per claim 2, Ma and Yang discloses the system of claim 1, wherein the computing system comprises at least one of
Furthermore, Ma discloses:
an orchestrator, a chat interface system, a human interface system, an information access device, a server, an AI/ML system, a cloud computing system, or a distributed computing system (A computing environment includes a chatbot server 131, Ma, para [0023]).
As per claim 3, Ma and Yang discloses the system of claim 1, wherein the communication session comprises one of
Furthermore, Ma discloses:
a chat session, a voice-only session, a telephone communication session, a video communication session, a multimedia communication session, a virtual reality ("VR")-based communication session, an augmented reality ("AR")-based communication session, or a mixed reality ("MR")-based communication session, wherein the NL input comprises corresponding one of NL text input, NL voice input, or NL sign language input (Entering the question may include typing the query into a user interface of a chatbot, such as the textbox 202, Ma, para [0080]).
As per claim 6, Ma and Yang discloses the system of claim 5, wherein
Furthermore, Yang discloses:
the similarity evaluation is performed using one of a similarity function, a similarity metric, or a neural network that has been trained on similar example pairs (d is a distance function or other function that computes a similarity between two inputs. For example, d may compute a Euclidean distance or a cosine similarity between the two token embeddings, Yang, para [0052])
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma and Yang by updating of response generation by an AI chatbot (Ma) with named entity recognition (Yang). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma and Yang in order to efficiently search and access user-specific data items (See Yang, para [0046])
As per claim 7, Ma and Yang discloses the system of claim 5, wherein
Furthermore, Yang discloses:
performing the similarity evaluation further comprises performing the similarity evaluation between the inputs of the dialogue context and inputs of dialogue contexts of at least one of a plurality of pairs of example dialogue contexts containing adversarial inputs and example outputs containing adversarial input mitigation responses, a plurality of pairs of example dialogue contexts containing off-topic inputs and example outputs containing off-topic input mitigation responses, or a plurality of pairs of example dialogue contexts containing non-adverse inputs and example outputs containing non-adverse responses, wherein the subset of example pairs further comprises at least one of a subset of the plurality of pairs of example dialogue contexts containing adversarial inputs and example outputs containing adversarial input mitigation responses, a subset of the plurality of pairs of example dialogue contexts containing off-topic inputs and example outputs containing off-topic input mitigation responses, or a subset of the plurality of pairs of example dialogue contexts containing non-adverse inputs and example outputs containing non-adverse responses (The support text may include one or more sentences that provide labelled examples and input text may include any text such as a sentence and it is desired to perform named entity recognition on the input text and input token embeddings 512 and support token embeddings 514 may be compared with each other and also nearest neighbor component 520 may compute distances between input token embeddings 512 and support token embeddings 514, Yang, para [0022], [0051]. Here, the labelled examples in the support text is analogous to the non-adverse inputs/non-adverse responses alternative)
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma and Yang by updating of response generation by an AI chatbot (Ma) with named entity recognition (Yang). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma and Yang in order to efficiently search and access user-specific data items (See Yang, para [0046])
As per claim 11, Ma discloses a computer-implemented method for implementing malicious input mitigation for artificial intelligence ("AI") and/or machine learning ("ML") models, the method comprising a computing system performing the following operations:
receiving, via a user interface, a natural language ("NL") input during a communication session for generating responses to NL inputs by one or more generative AI systems; (A system including a user interface, such as a computer, smart phone, tablet, etc., may allow a user to communicate with an artificial intelligence (AI) dialog generator, or chatbot, by entering query texts, Ma, para [0018]. A text box 202 of a chatbot may be presented where a query text may be entered by a user by typing on a keyboard or speaking into a microphone if the device is adapted with voice recognition capabilities and if query may be entered using natural language, Ma, para [0045]. One task of the NLP system is to generate natural responses to input queries using Natural Language Generation (NLG), herein referred to as dialog generation (DG), Ma, para [0002]. Here, the user interface is analogous to the computer and chat text box. The NL input aligns with the query text entered using natural language. The communication sessions are the user communicating with the AI dialog generator)
providing the second prompt to the second generative AI system, the second prompt causing the second AI/MIL-based system to determine an output by finding a probable continuation of the second prompt; (Receiving an input query by a variational autoencoder (VAE), embedding the input query into vector representations via the VAE, converting the vector representations into responses by a generative adversarial network (GAN), Ma, para [0089]. Here, the VSDG or the VAE-GAN is analogous to the second generative AI system and receiving the input query and converting vector representations into responses indicates providing the prompt.)
receiving the output; (Receiving, at a generator of the GAN (GAN-G), the response embedding from the decoder, and generating, with the GAN-G, a response sentence to the input query, Ma, para [0089]. Here, the received output is analogous to the generated response sentence produced by the GAN-G and then displayed by the chatbot system)
generating a response to the NL input based on the received output; and (The VAE may convert queries received by the chatbot to vectors that are then used by the neural network to generate a response, the GAN-G generating a response to the received response embeddings, Ma, para [0018]. The response to the NL input is analogous to chatbot/ GAN generated response sentence generated from vectors response embeddings derived from the user query)
causing, via the user interface, presentation of the generated response within the communication session (Displaying a response to the input query in a user interface. The response is displayed in a response dialog box 206 below the query text box 202. The answer to the question is displayed on a computer screen by the chatbot for the user to observe. Ma, para [0046], [0083], [0089]).
However, Ma does not explicitly disclose the limitation:
generating a first prompt, for a first generative AI system, comprising a dialogue context including the NL input;
providing the first prompt to the first generative AI system, the first prompt causing the first generative AI system to dynamically generate a subset of example pairs by filtering at least one of following example pairs based on results of similarity evaluation between inputs of the dialogue context and inputs of one or more of the following dialogue contexts: (a) a plurality of pairs of example dialogue contexts containing malicious inputs and example outputs containing malicious input mitigation responses; (b) a plurality of pairs of example dialogue contexts containing adversarial inputs and example outputs containing adversarial input mitigation responses; (c) a plurality of pairs of example dialogue contexts containing off-topic inputs and example outputs containing off-topic input mitigation responses; (d) a plurality of pairs of example dialogue contexts containing attack vector- based inputs and example outputs containing attack vector-based input mitigation responses; or (e) a plurality of pairs of example dialogue contexts containing non-adverse inputs and example outputs containing non-adverse responses;
receiving the dynamically generated subset of example pairs;
generating a second prompt, for a second generative AI system, comprising the dialogue context, including the NL input, and the subset of example pairs;
Yang discloses:
generating a first prompt, for a first generative AI system, comprising a dialogue context including the NL input; (The inputs to system 500 includes support text and input text, support token are obtained and the support token may be multiple sequences of support tokens corresponding to text, such as multiple sentences, Yang, para [0046]. Here, the prompt is interpreted as a model input package. The current dialogue context with the NL input is analogous to current input text/query text. The dynamically identified subset of example pairs aligns with the support text and examples selected by similarity)
providing the first prompt to the first generative AI system, the first prompt causing the first generative AI system to dynamically generate a subset of example pairs by filtering at least one of following example pairs based on results of similarity evaluation between inputs of the dialogue context and inputs of one or more of the following dialogue contexts: (a) a plurality of pairs of example dialogue contexts containing malicious inputs and example outputs containing malicious input mitigation responses; (b) a plurality of pairs of example dialogue contexts containing adversarial inputs and example outputs containing adversarial input mitigation responses; (c) a plurality of pairs of example dialogue contexts containing off-topic inputs and example outputs containing off-topic input mitigation responses; (d) a plurality of pairs of example dialogue contexts containing attack vector- based inputs and example outputs containing attack vector-based input mitigation responses; or (e) a plurality of pairs of example dialogue contexts containing non-adverse inputs and example outputs containing non-adverse responses; (The support text may include one or more sentences that provide labelled examples and the input text may include any text that may be input to a production NER system, Yang, para [0046]. Here, the labelled examples are interpreted as non-adverse support examples)
receiving the dynamically generated subset of example pairs; (A closest support token may be selected, and the first tag that is assigned to the first token may be determined from the tag of the closest support token, Yang, para [0088]. Here, the subset is analogous to the selected closest support token or example after similarity evaluation. Ma also discloses receiving, at a generator of the GAN (GAN-G), the response embedding from the decoder, and generating, with the GAN-G, a response sentence to the input query)
generating a second prompt, for a second generative AI system, comprising the dialogue context, including the NL input, and the subset of example pairs; (The inputs to system 500 includes support text and input text, Yang, para [0046]. Here, the model input package that include input text and support text is interpreted as the second prompt. The current chatbot query is analogous to the dialogue context including NL input. The subset example pairs maps to selected support examples. Ma also discloses that the query may be entered using natural language and the VSDG processes the query to form answers)
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma and Yang by updating of response generation by an AI chatbot (Ma) with named entity recognition (Yang). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma and Yang in order to efficiently search and access user-specific data items (See Yang, para [0046])
As per claim 13, Ma and Yang disclose the computer-implemented method of claim 11, wherein
Furthermore, Yang discloses:
the similarity evaluation is performed using one of a similarity function, a similarity metric, or a neural network that has been trained on similar example pairs (d is a distance function or other function that computes a similarity between two inputs. For example, d may compute a Euclidean distance or a cosine similarity between the two token embeddings, Yang, para [0052])
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma and Yang by updating of response generation by an AI chatbot (Ma) with named entity recognition (Yang). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma and Yang in order to efficiently search and access user-specific data items (See Yang, para [0046])
As per claim 15, Ma and Yang disclose the computer-implemented method of claim 14, wherein updating the set of example pairs comprises at least one of:
Furthermore, Yang discloses:
after becoming aware of new malicious inputs, prompting an AI/ML-based system to predict mitigation responses for the new malicious inputs, and adding the new malicious inputs and corresponding predicted mitigation responses to the subset of example pairs; after becoming aware of new adversarial inputs, prompting an AI/ML-based system to predict mitigation responses for the new adversarial inputs, and adding the new adversarial inputs and corresponding predicted mitigation responses to the subset of example pairs; after becoming aware of new off-topic inputs, prompting an AI/ML-based system to predict mitigation responses for the new off-topic inputs, and adding the new off- topic inputs and corresponding predicted mitigation responses to the subset of example pairs; after becoming aware of new attack vectors, prompting an AI/ML-based system to predict mitigation responses for the new attack vectors, and adding the new attack vectors and corresponding predicted mitigation responses to the subset of example pairs; or after becoming aware of new non-adverse inputs, prompting an AI/ML-based system to predict non-adverse responses for the new non-adverse inputs, and adding the new non-adverse inputs and corresponding predicted non-adverse responses to the subset of example pairs (The support text may include one or more sentences that provide labelled examples and input text may include any text such as a sentence and it is desired to perform named entity recognition on the input text and input token embeddings 512 and support token embeddings 514 may be compared with each other and also nearest neighbor component 520 may compute distances between input token embeddings 512 and support token embeddings 514, Yang, para [0022], [0051]. Here, the labelled examples in the support text is analogous to the non-adverse inputs/non-adverse responses alternative)
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma and Yang by updating of response generation by an AI chatbot (Ma) with named entity recognition (Yang). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma and Yang in order to efficiently search and access user-specific data items (See Yang, para [0046])
As per claim 16, Ma discloses a system for implementing malicious input mitigation for artificial intelligence ("AI") and/or machine learning ("ML") models, the system comprising: a computing system, comprising:
at least one processor; and
a computer storage medium communicatively coupled to the at least one processor, the computer storage medium storing instructions that, when executed by the at least one processor, causes the computing system to perform operations comprising:
receiving, via a user interface, a natural language ("NL") input during a communication session for generating responses to NL inputs by a generative AI model; (A system including a user interface, such as a computer, smart phone, tablet, etc., may allow a user to communicate with an artificial intelligence (AI) dialog generator, or chatbot, by entering query texts, Ma, para [0018]. A text box 202 of a chatbot may be presented where a query text may be entered by a user by typing on a keyboard or speaking into a microphone if the device is adapted with voice recognition capabilities and if query may be entered using natural language, Ma, para [0045]. One task of the NLP system is to generate natural responses to input queries using Natural Language Generation (NLG), herein referred to as dialog generation (DG), Ma, para [0002]. Here, the user interface is analogous to the computer and chat text box. The NL input aligns with the query text entered using natural language. The communication sessions are the user communicating with the AI dialog generator)
providing the prompt to the generative AI model, the prompt causing the AI model to generate an output; (Receiving an input query by a variational autoencoder (VAE), embedding the input query into vector representations via the VAE, converting the vector representations into responses by a generative adversarial network (GAN). Generating, with the GAN-G, a response sentence to the input query, Ma, para [0089]. Here, receiving the input query at the VAE and converting it into vector representations for response generation is analogous to providing the prompt)
receiving the output; (Receiving, at a generator of the GAN (GAN-G), the response embedding from the decoder, and generating, with the GAN-G, a response sentence to the input query, Ma, para [0089]. Here, the received output is analogous to the generated response sentence produced by the GAN-G and then displayed by the chatbot system)
generating a response to the NL input based on the received output; and (The VAE may convert queries received by the chatbot to vectors that are then used by the neural network to generate a response, the GAN-G generating a response to the received response embeddings, Ma, para [0018]. The response to the NL input is analogous to chatbot/ GAN generated response sentence generated from vectors response embeddings derived from the user query)
causing, via the user interface, presentation of the generated response within the communication session (Displaying a response to the input query in a user interface. The response is displayed in a response dialog box 206 below the query text box 202. The answer to the question is displayed on a computer screen by the chatbot for the user to observe. Ma, para [0046], [0083], [0089]).
However, Ma does not explicitly disclose the limitations:
evaluating a similarity between inputs of a current dialogue context and inputs of dialogue contexts of a plurality of pairs of example dialogue contexts and example outputs, wherein the plurality of pairs of example dialogue contexts and example outputs includes a plurality of first context response pairs containing adverse inputs and example outputs containing adverse input mitigation responses and a plurality of second context response pairs containing non-adverse inputs and example outputs containing proper non-adverse input responses;
dynamically generating a subset of similar example pairs of dialogue contexts and example outputs based on a subset of the plurality of first context response pairs and a subset of the plurality of second context response pairs as determined from results of the similarity evaluation;
receiving the dynamically generated subset of example pairs;
generating a prompt, for a generative AI model, comprising the dialogue context, including the NL input, and the dynamically generated subset of example pairs;
Yang discloses:
evaluating a similarity between inputs of a current dialogue context and inputs of dialogue contexts of a plurality of pairs of example dialogue contexts and example outputs, wherein the plurality of pairs of example dialogue contexts and example outputs includes a plurality of first context response pairs containing adverse inputs and example outputs containing adverse input mitigation responses and a plurality of second context response pairs containing non-adverse inputs and example outputs containing proper non-adverse input responses; (The support text may include one or more sentences that provide labelled examples and input text may include any text such as a sentence and it is desired to perform named entity recognition on the input text and input token embeddings 512 and support token embeddings 514 may be compared with each other and also nearest neighbor component 520 may compute distances between input token embeddings 512 and support token embeddings 514, Yang, para [0022], [0051]. Here, the labelled examples in the support text is analogous to the non-adverse inputs/non-adverse responses alternative)
dynamically generating a subset of similar example pairs of dialogue contexts and example outputs based on a subset of the plurality of first context response pairs and a subset of the plurality of second context response pairs as determined from results of the similarity evaluation; (Support text that provides a few examples compute an embedding vector for each word of the support text, compute an embedding vector for each word of input text and then computes distances between the embedding vectors. Where an input token and a support token have similar meanings or other similar aspects, the corresponding token embeddings are more likely to be close to each other, for each input token, determine which support token is closest to the input token and may compute a Euclidean distance or a cosine similarity, Yang, para [0052])
receiving the dynamically generated subset of example pairs; (Receive the sequence of input token embeddings and the sequence of support token embeddings, Yang, para [0051])
generating a prompt, for a generative AI model, comprising the dialogue context, including the NL input, and the dynamically generated subset of example pairs; (The inputs to system 500 includes support text and input text, support token are obtained and the support token may be multiple sequences of support tokens corresponding to text, such as multiple sentences, Yang, para [0046]. Here, the prompt is interpreted as a model input package. The current dialogue context with the NL input is analogous to current input text/query text. The dynamically identified subset of example pairs aligns with the support text and examples selected by similarity)
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma and Yang by updating of response generation by an AI chatbot (Ma) with named entity recognition (Yang). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma and Yang in order to efficiently search and access user-specific data items (See Yang, para [0046])
As per claim 17, Ma and Yang disclose the system of claim 16, wherein
Furthermore, Yang discloses:
the adverse inputs comprise at least one of malicious inputs, adversarial inputs, off-topic inputs, or attack vector-based inputs, wherein the adverse input mitigation responses comprise at least one of malicious input mitigation responses, adversarial input mitigation responses, off-topic input mitigation responses, or attack vector-based input mitigation responses (The support text may include one or more sentences that provide labelled examples and input text may include any text such as a sentence and it is desired to perform named entity recognition on the input text and input token embeddings 512 and support token embeddings 514 may be compared with each other and also nearest neighbor component 520 may compute distances between input token embeddings 512 and support token embeddings 514, Yang, para [0022], [0051]. Here, the labelled examples in the support text is analogous to the non-adverse inputs/non-adverse responses alternative)
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma and Yang by updating of response generation by an AI chatbot (Ma) with named entity recognition (Yang). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma and Yang in order to efficiently search and access user-specific data items (See Yang, para [0046])
Claims 4,5,8 and 12 are rejected under 35 U.S.C. 103 as being unpatentable over Ma et al. (US 20200226475 A1), hereinafter referred to as Ma, in view of Yang et al. (US 20210374347 A1), hereinafter referred to as Yang in further view of Bennett et al. (US 20080052077 A1), hereinafter referred to as Bennett.
As per claim 4, Ma and Yang discloses the system of claim 1, wherein
However, Ma in view of Yang does not explicitly disclose the limitation:
the dialogue context comprises at least one of a user profile associated with the user, a history of NL dialogue including user NL inputs and corresponding system responses, one or more action strings paired with their results, metadata extracted from previously retrieved data items, or other data associated with data items that were retrieved from the data storage system
Bennett discloses:
the dialogue context comprises at least one of a user profile associated with the user, a history of NL dialogue including user NL inputs and corresponding system responses, one or more action strings paired with their results, metadata extracted from previously retrieved data items, or other data associated with data items that were retrieved from the data storage system (The context or environment experienced by the user can be determined at any moment in time based at the selection made at the section level, Bennett, para [0199]. Here, the dialogue context is analogous to the context or environment experienced by the user. The section, product or page context used to determine which Q/A pairs are appropriate)
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma and Yang with Bennett by updating of response generation by an AI chatbot (Ma) and named entity recognition (Yang) with multi-language speech recognition (Bennett). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma and Yang with Bennett in order to determine a context containing the selection of the user (See Bennett, para [0199])
As per claim 5, Ma and Yang disclose the system of claim 1, wherein
However, Ma in view of Yang does not explicitly disclose the limitation:
dynamically identifying the subset of example pairs comprises performing a similarity evaluation between inputs of the dialogue context and inputs of dialogue contexts of the plurality of pairs of example dialogue contexts containing malicious inputs and example outputs containing malicious input mitigation responses, wherein the subset of example pairs comprises a subset of the plurality of pairs of example dialogue contexts containing malicious inputs and example outputs containing malicious input mitigation responses whose inputs are most similar to the inputs of the dialogue context
Bennett discloses:
dynamically identifying the subset of example pairs comprises performing a similarity evaluation between inputs of the dialogue context and inputs of dialogue contexts of the plurality of pairs of example dialogue contexts containing malicious inputs and example outputs containing malicious input mitigation responses, wherein the subset of example pairs comprises a subset of the plurality of pairs of example dialogue contexts containing malicious inputs and example outputs containing malicious input mitigation responses whose inputs are most similar to the inputs of the dialogue context (Only a limited subset of question-answer pairs 708 for example are appropriate for section 705, Bennett, para [0199]. The context appropriate question- answer pairs selected for the user’s current section are similar to the subset of example pairs)
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma and Yang with Bennett by updating of response generation by an AI chatbot (Ma) and named entity recognition (Yang) with multi-language speech recognition (Bennett). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma and Yang with Bennett in order to determine a context containing the selection of the user (See Bennett, para [0199])
As per claim 8, Ma and Yang discloses the system of claim 7, wherein
However, Ma in view of Yang does not explicitly disclose the limitation:
dynamically identifying the subset of example pairs comprises filtering at least one of the plurality of pairs of example dialogue contexts containing malicious inputs and example outputs containing malicious input mitigation responses, the plurality of pairs of example dialogue contexts containing adversarial inputs and example outputs containing adversarial input mitigation responses, the plurality of pairs of example dialogue contexts containing off-topic inputs and example outputs containing off-topic input mitigation responses, or the plurality of pairs of example dialogue contexts containing non- adverse inputs and example outputs containing non-adverse responses, based on results of the similarity evaluation
Bennett discloses:
dynamically identifying the subset of example pairs comprises filtering at least one of the plurality of pairs of example dialogue contexts containing malicious inputs and example outputs containing malicious input mitigation responses, the plurality of pairs of example dialogue contexts containing adversarial inputs and example outputs containing adversarial input mitigation responses, the plurality of pairs of example dialogue contexts containing off-topic inputs and example outputs containing off-topic input mitigation responses, or the plurality of pairs of example dialogue contexts containing non- adverse inputs and example outputs containing non-adverse responses, based on results of the similarity evaluation (The NLQS database 188 organization is intricately linked to the switched grammar architecture and only a limited subset of question-answer pairs 708 are appropriate for section 705, Bennett, para [0199]. Here, limiting the available full set of question-answer pairs to those appropriate in the current session based on similarity evaluation is analogous to filtering).
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma and Yang with Bennett by updating of response generation by an AI chatbot (Ma) and named entity recognition (Yang) with multi-language speech recognition (Bennett). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma and Yang with Bennett in order to determine a context containing the selection of the user (See Bennett, para [0199])
As per claim 12, Ma and Yang disclose the computer-implemented method of claim 11, wherein
However, Ma in view of Yang does not explicitly disclose the limitation:
the dialogue context comprises at least one of a user profile associated with the user, a history of NL dialogue including user NL inputs and corresponding system responses, one or more action strings paired with their results, metadata extracted from previously retrieved data items, or other data associated with data items that were retrieved from the data storage system
Bennett discloses:
the dialogue context comprises at least one of a user profile associated with the user, a history of NL dialogue including user NL inputs and corresponding system responses, one or more action strings paired with their results, metadata extracted from previously retrieved data items, or other data associated with data items that were retrieved from the data storage system (The context or environment experienced by the user can be determined at any moment in time based at the selection made at the section level, Bennett, para [0199]. Here, the dialogue context is analogous to the context or environment experienced by the user. The section, product or page context used to determine which Q/A pairs are appropriate)
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma and Yang with Bennett by updating of response generation by an AI chatbot (Ma) and named entity recognition (Yang) with multi-language speech recognition (Bennett). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma and Yang with Bennett in order to determine a context containing the selection of the user (See Bennett, para [0199])
Claims 9,10, 14, 18-20 are rejected under 35 U.S.C. 103 as being unpatentable over Ma et al. (US 20200226475 A1), hereinafter referred to as Ma, in view of Yang et al. (US 20210374347 A1), hereinafter referred to as Yang in further view of Bennett et al. (US 20080052077 A1), hereinafter referred to as Bennett in further view of Wysopal et al. (US 20120072968 A1), hereinafter referred to as Wysopal.
As per claim 9, Ma, Yang, Bennett disclose the system of claim 1, wherein the operations further comprise:
However, Ma, Yang, Bennett does not explicitly disclose the limitation:
updating a set of example pairs by adding at least one of one or more additional pairs of example dialogue contexts containing malicious inputs and example outputs containing malicious input mitigation responses
Wysopal discloses:
updating a set of example pairs by adding at least one of one or more additional pairs of example dialogue contexts containing malicious inputs and example outputs containing malicious input mitigation responses (The data, scripts and functions used to operate the various testing engines and the analysis engine 125 may be stored in a security-threat database 150. Portions of the threat database 150 may, in some cases, be provided by entities other than the entity operating the platform 105 on a subscription basis, allowing the database 150 to be kept up to date as threats and malware evolve over time, Wysopal, para [0054]. Here, the example pairs are interpreted as the security-threat database entries used by the testing engines and the malicious inputs are treated as threats and malware. The updating is analogous to keeping the threat database up to date as threats and malware evolve over time)
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma, Yang, Bennett with Wysopal by updating of response generation by an AI chatbot (Ma) and named entity recognition (Yang) and multi-language speech recognition (Bennett) with analysis of security flaws (Wysopal). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma, Yang, Bennett and Wysopal in order to update the database to be kept up to date as threats and malware evolve over time (See Wysopal, para [0054])
As per claim 10, Ma, Yang and Bennett discloses the system of claim 9, wherein updating the set of example pairs comprises:
However, Ma, Yang, Bennett does not explicitly disclose the limitation:
after becoming aware of new attack vectors, generating another prompt for the generative AI system to predict mitigation responses for the new attack vectors, and adding the new attack vectors and corresponding predicted mitigation responses to the subset of example pairs
Wysopal discloses:
after becoming aware of new attack vectors, generating another prompt for the generative AI system to predict mitigation responses for the new attack vectors, and adding the new attack vectors and corresponding predicted mitigation responses to the subset of example pairs (STEP 605: A new method for attacking XML interfaces is discovered and threat metadata M and general remediation information T are imported into the threat database. STEP 610: When a new attack vector is discovered, security researchers and developers create a new scan that detects instances of the vector. The new scan capability is classified, codified, and is added to the threat database 150, Wysopal, para [0089]. Here, when a new attack vector is discovered, general remediation information and new scan that detects the attack vector means that after becoming aware of new attack vectors, mitigation responses are predicted)
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma, Yang, Bennett with Wysopal by updating of response generation by an AI chatbot (Ma) and named entity recognition (Yang) and multi-language speech recognition (Bennett) with analysis of security flaws (Wysopal). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma, Yang, Bennett and Wysopal in order to update the database to be kept up to date as threats and malware evolve over time (See Wysopal, para [0054])
As per claim 14, Ma, Yang, Bennett disclose the computer-implemented method of claim 11, wherein the computing system further performs the following:
However, Ma, Yang, Bennett does not explicitly disclose the limitation:
updating a set of example pairs by adding at least one of: one or more additional pairs of example dialogue contexts containing malicious inputs and example outputs containing malicious input mitigation responses; one or more additional pairs of example dialogue contexts containing adversarial inputs and example outputs containing adversarial input mitigation responses; one or more additional pairs of example dialogue contexts containing off-topic inputs and example outputs containing off-topic input mitigation responses; one or more additional pairs of example dialogue contexts containing attack vector-based inputs and example outputs containing attack vector-based input mitigation responses; or one or more additional pairs of example dialogue contexts containing non-adverse inputs and example outputs containing non-adverse responses
Wysopal discloses:
updating a set of example pairs by adding at least one of: one or more additional pairs of example dialogue contexts containing malicious inputs and example outputs containing malicious input mitigation responses; one or more additional pairs of example dialogue contexts containing adversarial inputs and example outputs containing adversarial input mitigation responses; one or more additional pairs of example dialogue contexts containing off-topic inputs and example outputs containing off-topic input mitigation responses; one or more additional pairs of example dialogue contexts containing attack vector-based inputs and example outputs containing attack vector-based input mitigation responses; or one or more additional pairs of example dialogue contexts containing non-adverse inputs and example outputs containing non-adverse responses (Re-analysis is triggered by changes in the external environment (e.g., threat space, business intelligence, detected attacks) and/or the implementation of enhanced analysis capabilities. The relevant analysis workflows are updated with the new scan information and re-processed, Wysopal, para [0081]. Here, the threat space changes, detected attacks and new vulnerability classes are interpreted as malicious or adversarial inputs. The mitigation responses are analogous to the updated scan/remediation workflow information. Adding additional pairs is analogous to updating analysis workflows with new scan information)
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma, Yang, Bennett with Wysopal by updating of response generation by an AI chatbot (Ma) and named entity recognition (Yang) and multi-language speech recognition (Bennett) with analysis of security flaws (Wysopal). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma, Yang, Bennett and Wysopal in order to update the database to be kept up to date as threats and malware evolve over time (See Wysopal, para [0054])
As per claim 18, Ma, Yang, Bennett discloses the system of claim 16, wherein the operations further comprise:
However, Ma, Yang, Bennett does not explicitly disclose the limitation:
updating a set of example pairs by adding one or more additional pairs of example dialogue contexts containing adverse inputs and example outputs containing adverse input mitigation responses
Wysopal discloses:
updating a set of example pairs by adding one or more additional pairs of example dialogue contexts containing adverse inputs and example outputs containing adverse input mitigation responses (Define rescan conditions based on the application's attack profile. In this example, any new attack vectors against Java applications or XML would constitute a rescan condition. All virtual machines that include applications having metadata that includes both XML and web services are identified and the relevant analysis workflows are updated with the new scan information and reprocessed, Wysopal, para [0083]. The adverse inputs are analogous to new attack vectors and attack profile conditions. The adverse input mitigation responses are analogous to new scan information re-analysis/ remediation workflow updates. Updating the relevant security analysis workflows or threat database with new-adverse condition or remediation information is basically updating of an example-pair)
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma, Yang, Bennett with Wysopal by updating of response generation by an AI chatbot (Ma) and named entity recognition (Yang) and multi-language speech recognition (Bennett) with analysis of security flaws (Wysopal). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma, Yang, Bennett and Wysopal in order to update the database to be kept up to date as threats and malware evolve over time (See Wysopal, para [0054])
As per claim 19, Ma, Yang, Bennett discloses the system of claim 18, wherein updating the set of example pairs comprises:
However, Ma, Yang, Bennett does not explicitly disclose the limitation:
after becoming aware of new adverse inputs, prompting an AI/ML-based system to predict mitigation responses for the new adverse inputs, and adding the new adverse inputs and corresponding predicted mitigation responses to the subset of example pairs
Wysopal discloses:
after becoming aware of new adverse inputs, prompting an AI/ML-based system to predict mitigation responses for the new adverse inputs, and adding the new adverse inputs and corresponding predicted mitigation responses to the subset of example pairs (A decision to initiate a re-analysis can be based, for example, on a technological profile, metadata describing the functionality of a virtual machine and/or the applications operating within the machine, the deployment environment of the virtual machine, new information about vulnerabilities that may affect the virtual machine, and/or increases in a likelihood of a threat, Wysopal, para [0082]. STEP 605: A new method for attacking XML interfaces is discovered and threat metadata M and general remediation information T are imported into the threat database, Wysopal, para [0088]. Here, general remediation information T and scan workflow updates is interpreted to be the prediction mitigation responses)
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma, Yang, Bennett with Wysopal by updating of response generation by an AI chatbot (Ma) and named entity recognition (Yang) and multi-language speech recognition (Bennett) with analysis of security flaws (Wysopal). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma, Yang, Bennett and Wysopal in order to update the database to be kept up to date as threats and malware evolve over time (See Wysopal, para [0054])
As per claim 20, Ma, Yang, Bennett discloses the system of claim 16, wherein the operations further comprise:
However, Ma, Yang, Bennett does not explicitly disclose the limitation:
determining whether the dialogue context contains any adverse inputs; and
based on a determination that the dialogue context contains at least one adverse input, determining effectiveness of adverse input mitigation on the at least one adverse input;
wherein the generated response further comprises a message indicating presence of one or more adverse inputs and a message indicating effectiveness of adverse input mitigation on the at least one adverse input
Wysopal discloses:
determining whether the dialogue context contains any adverse inputs; and (Providing at least one test input and inspecting a response from the executing application to the at least one test input to detect one or more potential vulnerabilities. and inspecting a response from the executing application to the at least one test input to detect one or more potential vulnerabilities, Wysopal, claim 9. Here, determining whether the dialogue context contains adverse input aligns with providing test input and inspecting the application response to detect vulnerabilities and creating a security report and computing the score to another benchmark means determining effectiveness of mitigation)
based on a determination that the dialogue context contains at least one adverse input, determining effectiveness of adverse input mitigation on the at least one adverse input; (Creating a security report from the listing of potential vulnerabilities and computing a security score from the security report, Wysopal, claim 15)
wherein the generated response further comprises a message indicating presence of one or more adverse inputs and a message indicating effectiveness of adverse input mitigation on the at least one adverse input (Comparing the security score with at least one security score associated with an implementation of the computer system, Wysopal, claim 15)
A person of ordinary skill in the art before the effective filing date of the claimed invention would have combined Ma, Yang, Bennett with Wysopal by updating of response generation by an AI chatbot (Ma) and named entity recognition (Yang) and multi-language speech recognition (Bennett) with analysis of security flaws (Wysopal). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ma, Yang, Bennett and Wysopal in order to update the database to be kept up to date as threats and malware evolve over time (See Wysopal, para [0054])
Conclusion
Any inquiry concerning this communication or earlier communications from the
examiner should be directed to RAGHAVENDER CHOLLETI whose telephone number is (703) 756-1065. The examiner can normally be reached Monday - Thursday 8AM-5PM EST & Friday variable.
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, RUPAL DHARIA can be reached on (571) 272-3880. 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.
Respectfully Submitted
/RAGHAVENDER NMN CHOLLETI/Examiner, Art Unit 2492
/RUPAL DHARIA/Supervisory Patent Examiner, Art Unit 2492
Prosecution Insights