# Unified Detection of Digital and Physical Face Attacks

**Authors:** Debayan Deb, Xiaoming Liu, Anil K. Jain  
**arXiv:** [2104.02156](https://arxiv.org/abs/2104.02156)  
**Date:** 2021-04-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2104.02156v1/

## Key points

- State-of-the-art defense mechanisms have achieved near-perfect accuracies within specific attack categories
- These defenses often fail to generalize well across all attack categories
- UniFAD is a unified attack detection framework that addresses this limitation
- UniFAD automatically clusters 25 coherent attack types belonging to different categories
- Utilizes multi-task learning and k-means clustering for joint representations of coherent attacks
- Outperforms existing defense methods with a True Detection Rate (TDR) of 94.73% at a False Detection Rate (FDR) of 0.2%
- Achieved on a large dataset consisting of bona fide and attack images across all three categories
- Exceptional speed in detecting attacks, achieving detection within just 3 milliseconds on an Nvidia 2080Ti GPU
- High accuracy in identifying specific attack types (75.81%) and categorizing them into broader attack categories (97.37%)
- Presents a significant advancement in face attack detection by addressing the challenge of generalization across different attack categories

## Summary

In the field of face attack detection, state-of-the-art defense mechanisms have achieved near-perfect accuracies within specific attack categories, such as adversarial attacks, digital manipulation, or physical spoofs. However, these defenses often fail to generalize well when tested across all three categories. This lack of generalization can be attributed to the joint learning of incoherent attacks. To address this limitation, the authors propose a unified attack detection framework called UniFAD. This framework is capable of automatically clustering 25 coherent attack types that belong to the three aforementioned categories. UniFAD utilizes a multi-task learning approach combined with k-means clustering to learn joint representations for coherent attacks, while uncorrelated attack types are learned separately. The proposed UniFAD outperforms existing defense methods and their fusion techniques with an impressive True Detection Rate (TDR) of 94.73% at a False Detection Rate (FDR) of 0.2%. These results were obtained on a large dataset consisting of 341,000 bona fide images and 448,000 attack images across all three attack categories. Notably, UniFAD demonstrates exceptional speed in detecting attacks, achieving detection within just 3 milliseconds on an Nvidia 2080Ti GPU. Additionally, UniFAD exhibits high accuracy in identifying specific attack types (75.81%) and categorizing them into one of the three broader attack categories (97.37%). Overall, the proposed UniFAD framework presents a significant advancement in face attack detection by effectively addressing the challenge of generalization across different attack categories. Its superior performance and efficiency make it a promising solution for ensuring robust security against face attacks in various applications and scenarios.