Spoofing Detection

[DJERLO]

Addressing Spoofing Detection in Facial Recognition

As part of our attendance monitoring system, we recognize the critical need to implement effective spoofing detection measures to enhance the reliability and security of our facial recognition feature. Spoofing attacks, where unauthorized individuals attempt to gain access using photos, videos, or masks, pose significant risks to the integrity of our system.

To address this issue, we plan to utilize the dataset available at CASIA-FASD. This dataset contains a diverse range of spoofing attacks, including both 2D and 3D facial representations, which will enable us to train and evaluate our spoof detection algorithms effectively.

Objectives:

• Develop robust algorithms to differentiate between genuine and spoofed facial data.
• Implement real-time detection mechanisms to enhance user security during check-in/check-out processes.
• Minimize false acceptance rates to ensure only authorized individuals gain access.

By prioritizing spoofing detection, we aim to strengthen the overall security of our attendance monitoring system and build user trust in our facial recognition technology.

[DJERLO]

Here are Some Promising Result so far

Hardware Setup:

• Camera: A4TECH PK-635G Anti-glare USB Webcam with Built-in Microphone
• CPU: Intel® Core™ i3-9100F Processor
• GPU: GeForce GTX 1050 2G OC

Reference Image:
Hardware Setup Image

---

Model & Training Details:

• Model: We are using ResNet-50, a deep convolutional neural network (CNN), as the backbone for our anti-spoofing task.

• Pre-trained Dataset: The model was initially pre-trained on ImageNet1K, which contains over 1 million images across 1,000 classes, allowing the model to capture detailed feature representations.

• Fine-tuning Dataset: For anti-spoofing, we fine-tuned ResNet-50 using the CASIA-FASD dataset, which includes close-up images of both real and spoofed faces. This dataset contains various spoofing attacks such as printed photos, video replays, and 3D masks. You can find the dataset here on Kaggle.

• Purpose: ResNet-50 was trained to detect subtle differences between real and spoofed faces, using residual connections to enhance feature learning depth, crucial for distinguishing between genuine human faces and spoofing attempts like images, masks, or video replays.