This project aims to prevent corporate internal security problems (especially data leakage through storage device) using customized YOLOv3. You can find referenced code here.
Our code is not provided due to security concerns and it is Enterprise Cooperation Project.
Demo 1 - She doesn't belong to Insider class So, it can't detect her face.
Demo 2 - She belongs to Insider class So, it can detect her face.
The dataset is about 9,000 images included a total of 12 PCs and 5 people.
Using YOLO-v3, I used images which is someone inserts storage device (such as USB) into a PC.
Three out of five people were recognized through face recognition and we call these insider.
- We pre-trained the face recognition model to recognize the faces of these three people.
And the other two people, we call them outsider.
In the dataset, we cropped pixels (unfixed-sized) that someone inserts a storage device into a PC visibly. We defined this class as 'PC_Vis'.
If a device is inserted invisibly, We defined this class as 'PC_Invis.
Likewise, If a device is inserted into a laptop, We defined as 'Laptop_Vis' and 'Laptop_Invis'.
Finally, we defined a 'face' class when it detects human faces.
Face recognition was performed with YOLO-v3.
When a face is detected by YOLO-v3 (Does not distinguish who it is.), it determines whether it is an insider by calling a function of face recognition.
We just customized the YOLO-v3 model using only 5 classes (PC_Vis, PC_Invis, Laptop_Vis, Laptop_Invis, Face).
Additionally, a cam was connected to the Raspberry Pi for real-time detection.
As soon as abnormal behavior is detected, beeps have been generated and line message has been sent via Raspberry Pi.
YOLOv3: An Incremental Improvement by Joseph Redmon
You Only Look Once: Unified, Real-Time Object Detection by Joseph Redmon
Jungseob Lee / js-lee-AI / omanma1928@naver.com
Juhoon Kim / galaxy1014 / galaxy1014@naver.com
If you need advice on my project, feel free to email me.