Context-aware adaptive routing framework for resource-efficient deepfake-aware biometric authentication

Abstract

Deepfakes pose a significant risk to facial biometric identity verification systems. While modern detection models are precise, they require a lot of computing power. This can slow down real-time performance if used consistently for every login attempt. This paper suggests a context-aware adaptive authentication framework. It employs a lightweight Random Forest classifier to route login requests based on the estimated risk level. In this study, the deepfake detection is modelled with a simulation-based cost model rather than a fully implemented end-to-end detection system, allowing the proposed routing mechanism to be evaluated independently of any specific deepfake detection architecture. Experiments conducted on the Risk-Based Authentication (RBA) dataset demonstrate that the proposed approach achieves high security performance, with a Recall of 99.22% and a False Negative Rate (FNR) below 1%. Additionally, the method enhances system efficiency, reducing computing cost by 57.68%, energy consumption by 58.48%, and processing latency by 62.42% compared to an Always-heavy baseline. These results indicate that integrating context-aware risk assessment into biometric pipelines offers a practical solution for balancing strong security with resource efficiency in real-world deployments.