Smart Cloud Storage Using Face Recognition

Abstract -Data security and identity confirmation are criticalelements in modern digital society, given the rapidly-growingimpact that cloud computing has had on data storage.Unauthorized data access can cause severe privacy violationsand negatively impact businesses - often causing sensitivecustomer information to be lost or changed. This research paperpresents an advanced intelligent Smart Cloud Storage systemwith a highly effective Two-Factor Authentication (2FA) basedon a facial recognition pipeline. It achieves this by utilizing anintelligent fallback detection scheme based on multiplebackends, with the first level of detection having a fallbacklevel that prioritizes the RetinaFace detection system -providing greater reliability throughout this phase of theauthentication process. Depending on which biometric featuresets are being used, both FaceNet's and GhostFaceNet's use ofspecific distance metrics for feature validation guarantee highlevels of accuracy and context-awareness. All of theaforementioned characteristics will be combined in a new soft-voting ensemble architecture known as "Crop Once, EmbedEverywhere", allowing for both increased accuracy and speedwhen processing the full 2FA authentication pipeline.Theentire process has been validated using the Labeled Faces in theWild (LFW) dataset, generating performance benchmarks foreach architecture and experiment performed. Findings showthat FaceNet has an average identification accuracy of 98.85%at peak training time, while SFace has proven to be one of thefastest architectures with a PPS rating of 21.36 frames persecond (FPS)This approach provides an effective method forquickly comparing known data against new data whileproviding verification accuracy of 96.5% and having low errorrates of 3% during training verification tasks. In addition, inorder to provide the operational transparency requested by thecustomer, the system includes a visual diagnostic dashboardthat allows for real-time debugging of authentication failures,by providing structural similarity and embedding qualitycalculations