SecureResume: A Security-Aware AI/ML Framework for Robust LLM-Based Resume Screening

Abstract Large language models (LLMs) are increasingly used for first-pass resume screening, yet the resume is supplied by the party who benefits from a favourable outcome. Applicants can therefore conceal text from human reviewers (white or microscopic fonts, zero-width characters, homoglyphs) or embed instructions that steer the model . We present SecureResume, a layered framework that (i) parses documents in a render-aware manner and removes concealed or instruction-like content, (ii) scores per-object adversarial risk with an Isolation Forest, (iii) computes Sentence-BERT job–resume similarity, and (iv) forces an LLM to emit schema-validated claims that are verified against evidence in the sanitised text, producing an evidence-adjusted ranking score. We implement the system as a FastAPI/Next.js application and evaluate its security-aware parsing layer on a controlled benchmark of 400 synthetic PDF resumes covering nine concealment techniques. After fixing two defects we found in our own parser, the rule-based layer detects all seven concealment classes it targets (280/280) with no false positives on 40 clean resumes, but misses off-page and low-contrast text; the document-level risk score, as originally formulated, saturates and is not discriminative . We report these findings candidly and outline calibration and rendering-based extensions. The LLM verification stages are specified but not yet evaluated.