Scalable systems and methods for context-aware sensitive data detection, hierarchical labeling, and protection in natural language processing environments
The present disclosure relates to scalable systems and methods for detecting, labeling, and protecting sensitive data in natural language processing (NLP) environments. This includes NLP applications in artificial intelligence (AI) systems, such as language models (LMs) and generative AI (GenAI). More particularly, the present disclosure introduces a hierarchical, context-aware labeling mechanism that is optimized using an LM in conjunction with machine learning (ML) techniques to ensure the utility-preserving effective protection of sensitive data with, for example, minimal false positives and false negatives and/or optimal precision and recall (e.g., in terms of an F1 Score).
1 . A system, comprising:
a memory having, stored therein, computer program code; and
one or more processing units operatively coupled to the memory and configured to execute instructions in the computer program code that cause the one or more processing units to:
analyze unstructured textual data using at least one natural language processing (NLP) technique;
detect sensitive information in the unstructured textual data according to at least one defined policy;
assign tokens to portions of the unstructured textual data indicating sensitivity classifications, wherein the sensitivity classifications are based, at least in part, on determined context-driven meanings of the respective portions of unstructured textual data;
label sensitive content generated during AI inference-time output; and
assign token-based classification metadata to the labeled sensitive content post-AI inference-time.
2 . The system of claim 1 , wherein the portions of the unstructured textual data comprise at least one of the following: individual words; compound words;
word phrases; chunks of data; documents; or files.
3 . The system of claim 1 , wherein the tokens comprise metadata labels.
4 . The system of claim 1 , wherein the instructions in the computer program code further cause the one or more processing units to:
tokenize the detected sensitive information using reversible tokens configured to preserve semantic accuracy and utility during artificial intelligence (AI) inference-time usage.
5 . The system of claim 4 , wherein the reversible tokens are deterministic within a policy-defined context scope.
6 . The system of claim 4 , wherein the reversible tokens comprise temporally unique dynamically changing de-identifiers (DDIDs).
7 . The system of claim 1 , wherein the instructions in the computer program code further cause the one or more processing units to:
associate provenance metadata with at least some of the detected sensitive information, wherein the provenance metadata is configured to provide for enforcement of context-sensitive access controls to the respective sensitive information at AI inference-time output.
8 . The system of claim 1 , wherein the instructions in the computer program code further cause the one or more processing units to:
selectively reverse at least some of the tokenized post-AI inference-time content.
9 . The system of claim 8 , wherein the instructions in the computer program code further cause the one or more processing units to:
semantically rephrase at least some of the tokenized post-AI inference-time content.
10 . The system of claim 8 , wherein the instructions in the computer program code further cause the one or more processing units to:
tokenize at least some data inof the post-AI inference-time content.
11 . The system of claim 8 , wherein the instructions in the computer program code further cause the one or more processing units to:
enforce database query authorization by applying token-based classification metadata to at least some of the post-AI inference-time content.
12 . A non-transitory program storage device comprising instructions stored thereon to cause one or more processors to:
analyze unstructured textual data using at least one natural language processing (NLP) technique;
detect sensitive information in the unstructured textual data according to at least one defined policy; and
assign tokens to portions of the unstructured textual data indicating sensitivity classifications, wherein the sensitivity classifications are based, at least in part, on determined context-driven meanings of the respective portions of unstructured textual data; and
associate token-based provenance metadata with at least some of the detected sensitive information, wherein the token-based provenance metadata is configured to provide for enforcement of context-sensitive access controls to the respective sensitive information post-AI inference-time.
13 . The non-transitory program storage device of claim 12 , wherein the instructions further cause the one or more processors to:
tokenize the detected sensitive information using reversible tokens configured to preserve semantic accuracy and utility during AI inference-time usage.
14 . The non-transitory program storage device of claim 13 , wherein the reversible tokens are deterministic within a policy-defined context scope.
15 . The non-transitory program storage device of claim 12 , wherein the instructions further cause the one or more processors to:
label sensitive content generated during AI inference-time output; and
assign token-based classification metadata to the labeled sensitive content post-AI inference-time.
16 . The non-transitory program storage device of claim 15 , wherein the instructions further cause the one or more processors to:
selectively reverse at least some of the tokenized post-AI inference-time content.
17 . The non-transitory program storage device of claim 16 , wherein the instructions further cause the one or more processors to:
semantically rephrase at least some of the tokenized post-AI inference-time content.
18 . The non-transitory program storage device of claim 16 , wherein the instructions further cause the one or more processors to:
tokenize at least some of the post-AI inference-time content.
19 . The non-transitory program storage device of claim 12 , wherein the instructions further cause the one or more processors to:
enforce database query authorization by applying token-based classification metadata to at least some of the post-AI inference-time content.
20 . A computer-implemented method, comprising:
analyzing unstructured textual data using at least one natural language processing (NLP) technique;
detecting sensitive information in the unstructured textual data according to at least one defined policy;
assigning tokens to portions of the unstructured textual data indicating sensitivity classifications, wherein the sensitivity classifications are based, at least in part, on determined context-driven meanings of the respective portions of unstructured textual data;
labeling sensitive content generated during AI inference-time output; and
assigning token-based classification metadata to the labeled sensitive content post-AI inference-time.
21 . The computer-implemented method of claim 20 , further comprising:
tokenizing the detected sensitive information using reversible tokens configured to preserve semantic accuracy and utility during inference-time usage.
22 . The computer-implemented method of claim 21 , wherein the reversible tokens comprise temporally unique dynamically changing de-identifiers (DDIDs).
23 . The computer-implemented method of claim 22 , further comprising:
associating provenance metadata with at least some of the detected sensitive information, wherein the provenance metadata is configured to provide for enforcement of context-sensitive access controls to the respective sensitive information at AI inference-time output.
24 . The computer-implemented method of claim 21 , further comprising:
selectively reversing at least some of the tokenized post-AI inference-time content.
25 . The computer-implemented method of claim 24 , further comprising:
semantically rephrasing at least some of the tokenized post-AI inference-time content.
26 . The computer-implemented method of claim 24 , further comprising:
tokenizing at least some of the post-AI inference-time content.
27 . The computer-implemented method of claim 20 , further comprising:
enforcing database query authorization by applying token-based classification metadata to at least some of the post-AI inference-time content.