PROXY INTERVIEW IDENTIFICATION AND VERIFICATION

Abstract

Remote interviews provide flexibility but create a persistent identity-integrity problem: the person who registers may not be the person who remains on camera and speaks throughout the session. This paper presents a fully offline Proxy Interview Identification and Verification system that continuously evaluates the candidate through face recognition, voice consistency, face-presence monitoring, and liveness analysis. During registration, 3-8 webcam photographs are used to train a candidate-specific Local Binary Patterns Histogram model, while a recorded voice sample is converted into a Mel-Frequency Cepstral Coefficient fingerprint. During an interview, browser-captured frames are analysed with OpenCV Haar Cascades and LBPH matching; audio snippets are compared using MFCC vectors and cosine similarity; and texture sharpness plus motion variance provide explainable liveness evidence. The Flask backend stores candidates, sessions, and timestamped events in SQLite and calculates a final Identity Confidence Score using documented weights of 50% face, 30% voice, and 20% liveness. The resulting verdict is Verified, Needs Review, or Failed. Real-time alerts identify no-face, multiple-face, face-mismatch, voice-mismatch, and liveness-failure events, while the administrator dashboard preserves a complete session timeline. Functional verification in the project confirms biometric enrollment, continuous monitoring, alert generation, score aggregation, protected administration, and offline execution using open-source libraries without external APIs or downloaded models. To illustrate the behaviour of the configured fusion rule, this paper additionally presents an eight-session simulated result set; these dummy values are clearly labelled and are not claimed as measured biometric accuracy.