AI Shield: A Privacy-Preserving Hybrid System for Explainable Scam and Phishing Risk Assessment

AI Shield is a privacy-preserving, explainable scam and phishing risk-assessment system that combines a bundled multinomial Naive Bayes model, hand-authored security rules, and behavioral indicators to generate a 0–100 risk score with SAFE, SUSPICIOUS, or SCAM explanations.This work presents the system architecture, privacy-focused browser workflow, and external transfer evaluations using the Nazario_5 phishing corpus and UCI SMS Spam Collection. On Nazario_5, AI Shield achieved 77.31% precision and 96.73% specificity, but 10.67% recall. On UCI SMS Spam, recall was 0.40%, demonstrating substantial limitations under domain shift.The findings characterize the current prototype as a conservative, high-specificity warning layer rather than a production-ready phishing detector or general spam classifier. The paper discusses limitations, ethical and privacy considerations, reproducibility, and directions for improving phishing-specific training, multilingual detection, URL analysis, visual deception detection, and robustness against adversarial obfuscation.