A Hybrid AI Framework for Privacy Risk Analysis in Digital Footprints Using NLP, Machine Learning, a

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 05 | May 2026 www.irjet.net p-ISSN: 2395-0072

A Hybrid AI Framework for Privacy Risk Analysis in Digital Footprints Using NLP, Machine Learning, and Large Language Models

Shaik Sumaiya Sultana¹, Palle Sahasra², Jade Vaishnavi³, Vyahruti Cheemalakonda´, Ms. K. Priyanka

¹²³´ Department of Computer Science and Engineering, Keshav Memorial Institute of Technology, Hyderabad, India

Abstract - The rapid growth of online platforms has increased the risk of unintentional disclosure of sensitive personal information. This paper proposes an AI-powered PrivacyFootprintAnalyzerthatdetectsandevaluatesprivacy risks in user-generated text. The system integrates Natural Language Processing (NLP) for identifying personally identifiable information, a Machine Learning model for risk classification, and a Large Language Model (LLM) for generatingcontext-awarerecommendationsandprivacy-safe text rewriting. A hybrid architecture combining detection, scoring,andrecommendationmodulesisimplementedwithina unified pipeline. Experimental results demonstrate high accuracy and low response time, enabling proactive privacy protection.Theproposedsystemenhancesuserawarenessand supports safer digital content sharing.

Key Words: Digital Footprint, Privacy Risk Analysis, Natural Language Processing, Machine Learning, Large Language Models, Cybersecurity

1. INTRODUCTION

The rapid proliferation of digital platforms has significantly transformed communication and information sharing,resultingintheexpansionofusers'digitalfootprints. These footprints often contain sensitive personal information, making individuals vulnerable to privacy breachesandcyberthreats[1].

Usersfrequentlypostcontentwithoutrecognizing implicit disclosures such as phone numbers, addresses, affiliations,orbehavioralpatterns.Theseexposurescanbe exploited for malicious purposes including social engineering,impersonation,andtargetedattacks[3].

While traditional cybersecurity tools focus on securing systems and credentials, they fail to address prepublication privacy risks in user-generated content. This createsacriticalgapinproactiveprivacyprotection.

Toaddressthischallenge,thispaperpresentsanAIdrivenPrivacyFootprintAnalyzerthatevaluatestextualdata before it is shared. The system leverages a hybrid architecture combining NLP for detection, ML for classification,andLargeLanguageModels(LLMs)[8],[9]for contextualreasoningandrecommendationgeneration.

KeyContributions

● AhybridAIpipelineintegratingNLP,ML,and LLMs

● Amulti-factorprivacyriskscoringmodel

● Context-awarerecommendationandtext rewritingsystem

● End-to-endimplementationwithreal-timeuser interaction

2. RELATED WORK

Earlyprivacydetectionsystemsreliedonrule-based approaches, including regular expressions and keyword matching.Whileefficientforstructureddata,thesemethods lackcontextualunderstandingandfailindynamicreal-world scenarios.

NLP-based systems improved detection through NamedEntityRecognition(NER)[1],enablingidentification ofentitiessuchasnames,locations,andorganizations.Tools like Microsoft Presidio [4] and spaCy [5] provide scalable solutions but primarily focus on extraction rather than interpretation.

Machine learning approaches introduced risk classification[6],leveragingfeaturessuchasentityfrequency andtextcomplexity.However,thesemodelsdependheavily onfeatureengineeringandlacksemanticreasoning.

Recent advancements in Large Language Models (LLMs) [8], [9] enable contextual understanding and generationofhuman-likeresponses.LLMsenhanceprivacy systems by providing explanations and actionable recommendations.

Despitetheseadvancements,existingsystemssufferfrom:

● Lackofunifiedarchitecture

● Absenceofexplainableriskscoring

● Limiteduser-centricrecommendations

● Fragmentedworkflows

Theproposedsystemaddressesthesegapsthrougha fullyintegratedandintelligentpipeline.

3. PROBLEM STATEMENT

Despiteincreasingawarenessofcybersecurity,users continuetoexposesensitiveinformationdueto:

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 05 | May 2026 www.irjet.net p-ISSN: 2395-0072

 Lackofreal-timefeedbackbeforesharingcontent

 Absence of contextual understanding in existing tools

 No unified system for detection, scoring, and mitigation

 Limitedaccessibilityfornon-technicalusers

There is a need for a system that can proactively analyze textual data, quantify privacy risks, and provide actionableinsightsinasingleworkflow.

4. PROPOSED SYSTEM

The proposed system follows a layered hybrid architecture designed for accuracy, scalability, and interpretability.

Fig 1: Sequence Diagram of Privacy Footprint Analyzer

ThefunctionalinteractionsofthesystemareillustratedinFig 1.

1) 4.1 NLP Layer (PII Detection)

● UtilizesspaCy[5]andMicrosoftPresidio[4] Detects:

○ Names

○ Emails

○ Phonenumbers

○ Locations

○ Organizations

4.2 Feature Engineering Layer

Extractedfeaturesinclude:

● Entitycount

● Entitydensity

● Sensitivekeywordfrequency

● Textlength

● Contextualindicators

4.3 Machine Learning Layer

● Model:RandomForestClassifier[6]

● Output:

○ RiskScore(0–100)

○ RiskCategory(Low/Medium/High)

4.4 LLM Layer (Intelligent Assistance)

● IntegratedusingLangChainagents[10]

● Generates:

○ Context-awarerecommendations

○ Privacy-saferewrittentext

○ Explanationsofdetectedrisks

4.5 System Interface

● Frontend:React-baseddashboard

● Backend:FastAPI

● Database:MongoDB

5. METHODOLOGY

The system follows a structured pipeline. The system workflowisillustratedinFig2.

1. Step1:InputAcquisition

UserprovidestextviaUIorfileupload.

2. Step2:EntityDetection

NLPmodelsextractPIIentitieswithcontextualtagging.

3. Step3:FeatureExtraction

QuantitativefeaturesarecomputedforMLprocessing.

4. Step4:RiskScoringModel

Theprivacyriskscoreiscalculatedas: R=(S×D×F)/C

Where:

● S=Sensitivityofdetectedentities

● D=Entitydensity

● F=Frequencyofoccurrence

● C=Contextualprotectionfactor

Theapproachisinspiredbystandardriskmodeling techniques[3].

5. Step5:RiskClassification

RandomForestclassifiesriskintodiscretecategories.

6. Step6:LLM-BasedRecommendation

LLMgenerates:

● Riskexplanations

● Mitigationsuggestions

● Saferewrittencontent

Step7:Visualization

Resultsdisplayedwith:

● Riskscoreindicators

● Highlightedentities

● Recommendationpanels

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 05 | May 2026 www.irjet.net p-ISSN: 2395-0072

Fig 2: Sequence Diagram of Privacy Footprint Analyzer

Fig 2 illustrates the interaction flow of the system. The process begins with user input, followed by preprocessing and PII detection using NLP techniques. The extracted featuresarepassedtothemachinelearningmodelforrisk scoringandclassification.TheLLMmodulethengenerates recommendationsandrewrittentext,whicharedisplayedto theuserthroughtheinterface.

6. SYSTEM ARCHITECTURE

The system follows modular client-server architecture:

● UserInterface(React)

● APILayer(FastAPI)

● AIProcessingPipeline

● DatabaseLayer(MongoDB)

TheoverallsystemdesignisshowninFig3.

Fig3showstheoverallarchitectureofthesystem.Itconsists of a user interface, API layer, AI processing pipeline, and database layer. The modular design ensures efficient data flow, scalability, and integration of NLP, ML, and LLM components. Thisdesignensuresscalability,modularity,andefficientdata flow.

7. RESULTS AND DISCUSSION

Theproposedsystemwasevaluatedusingstandard performance metrics [3] to assess its effectiveness in

detectingandclassifyingprivacyrisks.Thesystemoutputs areshowninFig4,andperformanceevaluationispresented inFig5.

PerformanceMetrics

Table 1: Performance Evaluation Metrics of the Proposed System

8. KeyObservations

● HighaccuracyindetectingPIIacrossvariedinputs

● Effectiveclassificationofprivacyrisklevels

● LLM-generatedrecommendationssignificantly improvedusability

● Reducedfalsepositivescomparedtorule-based systems

Fig 4: User Interface Showing Privacy Risk Detection and Recommendations

Fig 4 presents the system output displayed to the user. It highlightsdetectedsensitiveentities,calculatedrisk score, andAI-generatedrecommendationsalongwithaprivacy-safe rewrittenversionoftheinputtext.

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056

Volume: 13 Issue: 05 | May 2026 www.irjet.net p-ISSN: 2395-0072

Fig5showsthesystemoutputforaGitHubprofileanalysis.It displaysthecalculatedprivacyriskscore,detectedsensitive informationsuchasemailexposure,anduserprofiledetails. The interface provides a clear visualization of risk levels alongwithcategorizedinsightsforimproveduserawareness.

Fig6illustratestheperformanceevaluationoftheproposed systemusingmetricssuchasaccuracy,precision,recall,and F1-score.Theresultsindicatehigheffectivenessindetecting andclassifyingprivacyrisks. Comparative Insight

Comparedtotraditionaltools:

Table 2 presentstheperformanceevaluationofthe proposedsystem. Feature Traditional Tools Proposed System

Context

RiskScoring

Recommendation s None AI-generated

Usability Complex User-friendly

8. CONCLUSIONS

This paper presents a comprehensive AI-driven solutionforanalyzingandmitigatingprivacyrisksintextual data. By integrating NLP, ML, and LLM technologies, the system provides accurate detection, meaningful risk assessment,andactionablerecommendations.

The proposed system shifts privacy protection from a reactive approach to a proactive, user-centric model [3], thereby enabling safer digital interactions and improved privacyawarenessamongusers.

Theexperimentalresultsvalidatetheeffectiveness and practicality of the proposed approach in real-world scenarios.

REFERENCES

[1] D. Jurafsky and J. H. Martin, Speech and Language Processing,3rded.,Pearson,2023.

[2] J. Devlin et al., “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,” arXiv preprint arXiv: 1810.04805, 2019.

[3] T. Fawcett, “An Introduction to ROC Analysis,” Pattern Recognition Letters,2006.

[4] Microsoft, “Presidio: Data Protection and PII Detection,”2023.

[5] Explosion AI, “spaCy: Industrial-Strength NLP Library,”2023.

[6] Scikit-learn,“MachineLearninginPython,”2023.

[7] L.Tunstalletal., Natural Language Processing with Transformers,O’Reilly,2022.

[8] OpenAI,“GPT-4TechnicalReport,”2023.

[9] MetaAI,“LLaMA3:OpenFoundationModels,”2024.

[10] LangChain,“LangChainDocumentation,”2023.

[11] MongoDBInc.,“MongoDBDocumentation,”2023.

[12] S.Ramírez,“FastAPIDocumentation,”2023.