CogniSentinel :A Smart Tool to Predict Crime Using Face and Web Data

Abstract

Abstract— Crime patterns have shifted heavily toward digital platforms in the last decade. Traditional policing methods that respond after an offense has occurred often fail to prevent harm. This paper presents a multimodal deep learning framework that merges web browsing behavior, facial expression analysis, and geolocation metadata to estimate criminal intent in real time. The proposed architecture runs two models side by side. A three
dimensional convolutional neural network handles temporal facial emotion data while a long short-term memory network captures sequential web and location patterns. Their individual risk scores pass through a late fusion algorithm that assigns a weighted final score. A built-in legal validation layer checks for proper authorization before any data enters the pipeline. Experiments on one thousand simulated user profiles show that the combined
approach reaches 94.5 percent accuracy, outperforming single modality baselines by roughly fifteen percentage points. Precision stands at 0.92 and the whole pipeline runs under one second on consumer-grade hardware. A risk stratification module sorts output into green, yellow, and red tiers so that a human analyst always
reviews high-risk flags before any action is taken. The study also discusses ethical safeguards, cultural bias risks, and possible future additions such as voice stress detection and blockchain-secured audit logs.