> ## Documentation Index > Fetch the complete documentation index at: https://docs.oloid-demo.com/llms.txt > Use this file to discover all available pages before exploring further. # Machine learning model details - Windows Login - Gadget detection > Model Used : Primary model, Secondary Model Model Used : Primary model, Secondary Model ## Overview Oloid has developed an advanced gadget detection feature as part of its Windows Login solution, leveraging computer vision and deep learning techniques to enhance security and user experience. At the core of this system is a customized implementation of the Primary Model object detection model, optimized for identifying unauthorized or unusual devices during login events. ## Key Objectives * Detect presence of external gadgets such as mobile phones, tablets, or unauthorized USB devices during Windows login. * Enhance workplace security by flagging potential security breaches. * Automate visual gadget detection to reduce reliance on manual surveillance. **Model Selection – Oloid’s Primary Model** has following features: * High inference speed and real-time processing capabilities. * Accuracy in object detection with fewer false positives. * Lightweight architecture suitable for on-device and edge deployments. * Robust support for custom object classes and transfer learning. **Presence Detection Prediction values for different scenarios:** **For Model**: oloid-nano-tiny-512.onnx **Windows Version**: Windows 11 pro Camera: Integrated System Camera Range of CPU usage during normal time CPU Spike during alert transaction creation Prediction of person/multi persons/Gadget Single Face 4 - 8 \<= 8 Clear face Infront of camera with or without specs or mask \>0.91 If Person is >=2 meters away \>0.80 to \<0.89 Person with specs also placed hand on face 0.80 -0.89 Note – If person at same distance in front of camera moves little bit continuously the values are coming in the range of 0.80 - 0.91 No Face If person not present in front of camera alert transactions are creating properly as per the alert time. Multi Face 4 - 8 \<= 8 People are side by side with clear face Infront of camera with or without specs or mask 1st person \>0.90 2nd person \>0.90 People are back-to-back with clear face Infront of camera with or without specs or mask 1st person \>0.90 2nd person 0.81 – 0.89 If 1st Person is in front of camera and 2nd person is >=2 meters away 1st person \>0.90 2nd person 0.51 – 0.75 Note – If 2 persons at same distance in front of camera back-to-back or side by side moves little bit continuously the values are coming in the range i.1st person – 0.85 – 0.94 ii. 2nd person – 0.75 – 0.85 Gadget 4 - 8 \<= 8 Gadget near to camera 0.65 – 0.81 Gadget near to person ear facing camera Lense towards camera 0.41 – 0.65 Gadget near to person ear facing camera Lense out of camera 0.21 – 0.45 Note – If person at same distance in front of camera place gadget in front of him (eg: using mobile with little face cover with mobile) with Lense little down the values are coming in the range of 1.Person – 0.81 – 0.89 2.Gadget – 0.45 – 0.61 For Model: oloid-small-mini-512.onnx Windows Version: Windows 11 pro Camera: Integrated System Camera Range of CPU usage during normal time CPU Spike during alert transaction creation Prediction of person/multi persons/Gadget Single Face 5 - 9 \>= 9 Clear face Infront of camera with or without specs or mask \>0.92 If Person is =2 meters away \>0.80 to \<0.90 Person with specs also placed hand on face \>0.91 Note – If person at same distance in front of camera moves little bit continuously the values are coming in the range of >0.91 No Face If person not present in front of camera alert transactions are creating properly as per the alert time. Multi Face 5 - 9 \>10 People are side by side with clear face Infront of camera with or without specs or mask 1st person \>0.90 2nd person \>0.90 People are back-to-back with clear face Infront of camera with or without specs or mask 1st person \>0.91 2nd person 0.81 – 0.91 If 1st Person is in front of camera and 2nd person is >=2 meters away 1st person \>0.90 2nd person 0.61 – 0.80 Note – If 2 persons at same distance in front of camera back-to-back or side by side moves little bit continuously the values are coming in the range i.1st person – >0.91 ii. 2nd person – 0.81 – 0.90 Gadget 5 - 9 \> 10 Gadget near to camera 0.65 – 0.81 Gadget near to person ear facing camera Lense towards camera 0.51 – 0.75 Gadget near to person ear facing camera Lense out of camera 0.21 – 0.45 Note – If person at same distance in front of camera place gadget in front of him (eg: using mobile with little face cover with mobile) with Lense little down the values are coming in the range of 1.Person – 0.81 – 0.91 2.Gadget – 0.61 – 0.75 Model Architecture & Training: * Training Dataset: Curated dataset containing images from office environments with labeled gadgets (phones, laptops, USBs, tablets, etc.). * Data Augmentation: Techniques such as random cropping, brightness variation, rotation, and occlusion applied to improve generalization. * Transfer Learning: Pretrained weights used for faster convergence and improved accuracy. Detection Pipeline: 1. Image Capture: Webcam or external camera captures login environment. 2. Preprocessing: Image resized and normalized to match model input specifications. 3. Inference: Oloid’s models detect and classifies objects in real-time. 4. Post-processing: Bounding boxes filtered based on confidence scores and non-max suppression. 5. **Alerting & Logging:** If unauthorized gadgets are detected, alerts are triggered and the event is logged. **Integration with Windows Login:** * The detection module is embedded within the Oloid login agent. * Before login is authorized, the gadget detection module performs a real-time scan. * Based on detection results, login can be approved, delayed for review, or blocked. **Performance Metrics:** * **Detection Accuracy:** \~92% mAP on internal validation set. * **Latency:** Average inference time \< 100ms per frame on edge GPU. * **False Positive Rate:** Maintained under 3% with ongoing model tuning. **Security and Privacy Considerations:** * All detection is done locally to preserve privacy. * No images are transmitted or stored unless flagged for review with user consent. * Model updates are securely deployed over the air. **Conclusion:** Oloid's gadget detection feature for Windows Login provides a proactive layer of security in enterprise environments. By combining real-time object detection with smart login workflows, it helps ensure secure and compliant access to Windows systems.