# Where Do Deep Fakes Look? Synthetic Face Detection via Gaze Tracking

**Authors:** Ilke Demir, Umur A. Ciftci  
**arXiv:** [2101.01165](https://arxiv.org/abs/2101.01165)  
**Date:** 2021-01-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2101.01165v2/

## Key points

- Authors Ilke Demir and Umur A. Ciftci address the concern of deep fake technology's impact on society
- Focus on biological signals for detecting authenticity signatures in videos before manipulation
- Novel approach identifies distinct eye and gaze features in deep fakes compared to authentic videos
- Features compiled into signatures for analysis based on geometric, visual, metric, temporal, and spectral variations
- Development of a deep neural network model for classifying videos as real or fake with high accuracy rates
- Model achieves impressive accuracy rates on various deep fake datasets
- Outperforms existing detectors without incorporating proposed gaze signatures
- Conducted ablation studies to explore different features, architectures, sequence durations, and post-processing artifacts in deep fake detection
- Study set to appear in ACM ETRA 2021 proceedings and contributes significantly to addressing challenges posed by synthetic media manipulation techniques

## Summary

In their paper titled "Where Do Deep Fakes Look? Synthetic Face Detection via Gaze Tracking," authors Ilke Demir and Umur A. Ciftci address the growing concern surrounding deep fake technology and its potential impact on society. With the democratization of AI, deep fake generators have become more accessible, leading to dystopian scenarios that erode trust in social interactions. The authors focus on a specific domain, namely biological signals, which has garnered attention for its potential in detecting authenticity signatures in real videos before they are manipulated by generative algorithms. The researchers propose a novel approach by identifying distinct eye and gaze features exhibited by deep fakes compared to authentic videos. They compile these features into signatures and conduct a detailed analysis to differentiate between real and fake videos based on geometric, visual, metric, temporal, and spectral variations. By leveraging this information, the authors develop a deep neural network model capable of classifying any video as either real or fake when encountered in the wild. To validate their method, the authors evaluate their approach using various deep fake datasets. Impressively, their model achieves high accuracy rates of 92.48% on FaceForensics++, 80.0% on Deep Fakes (in the wild), 88.35% on CelebDF, and an impressive 99.27% on DeeperForensics datasets. This outperformance extends to existing deep and biological fake detectors with complex network architectures even without incorporating the proposed gaze signatures. Furthermore, the researchers conduct ablation studies to explore different features, architectures, sequence durations, and post-processing artifacts involved in deep fake detection. Through their comprehensive analysis and experimentation . This study is set to appear in the proceedings of ACM ETRA 2021 and contributes significantly to ongoing efforts in addressing the challenges posed by synthetic media manipulation techniques like deep fakes.