Evaluation of Multimodal Sensitive Data Detection and Redaction in an Offline Desktop Application

The increasing use of digital documents, images, and audio recordings creates challenges in protecting personally identifiable information (PII). This paper presents KavachAI, an offline-oriented desktop application designed to detect and redact sensitiveinformation from text, images, and audio-derived transcripts. The system combines regular-expression matching and named entity recognition (NER) for text analysis, optical character recognition (OCR) for extracting text from images, and speech-to-text transcription for audio processing. Extracted text is analyzed using the text detection pipeline to identify sensitive information and generate masked output. The application is designed to minimize cloud dependency by processing supported inputslocally. Its evaluation considers detection performance and redaction effectiveness across the three input modalities. Preliminary testing indicates that text detection and transcript masking work for the tested samples, while image processing exhibits extraction and redaction limitations. Audio transcription and transcript masking function in the tested workflow, although audio beep functionality remains incomplete. These findings demonstrate the feasibility of a unified multimodal privacy-protection workflow while highlighting the need for improved document recognition, fewer false detections, and more reliable redaction.