Image anonymization is the practice of modifying or removing sensitive information from images to protect privacy. While important for compliance with privacy regulations, anonymization often reduces data quality, making it more difficult for machine vision to develop. There are several challenges, such as data degradation, balancing privacy and utility, creating efficient algorithms, and negotiating moral and legal issues. A proper compromise must be made to ensure privacy while improving machine vision research and applications.
Previous approaches to image anonymization include traditional methods such as blurring, masking, encryption, and bundling. Recent work focuses on realistic anonymization using generative models to replace identities. However, many methods lack formal guarantees of anonymity and other cues in the image can still reveal identity. Limited studies have explored the impact on computer vision models, with variable effects depending on the task. Public anonymous data sets are rare.
In recent research, researchers at the Norwegian University of Science and Technology have turned their attention to crucial computer vision tasks in the context of autonomous vehicles, specifically instance segmentation and human pose estimation. They have evaluated the performance of the full-body and face anonymization models implemented in DeepPrivacy2, with the aim of comparing the effectiveness of realistic anonymization approaches with conventional methods.
The steps proposed to evaluate the impact of anonymization by the article are the following:
- Anonymization of common computer vision data sets.
- Training of various models using anonymized data.
- Evaluation of the models on the original validation data sets
The authors propose three full-body and face anonymization techniques: blurring, masking, and realistic anonymization. They define the anonymization region based on the instance segmentation annotations. Traditional methods include Gaussian masking and blurring, while realistic anonymization uses pre-trained models from DeepPrivacy2. The authors also address global context issues in full-body synthesis through histogram equalization and latent optimization.
The authors conducted experiments to evaluate models trained on anonymous data using three data sets: COCO Posing Estimation, Cityscape Instance Segmentation, and BDD100K Instance Segmentation. Face anonymization techniques did not show significant performance differences in the Cityscapes and BDD100K data sets. However, for COCO pose estimation, both the masking and blur techniques led to a significant drop in performance due to the correlation between blur/masking artifacts and the human body. Full-body anonymization, whether traditional or realistic, resulted in decreased performance compared to the original data sets. Realistic anonymization performed better, but still degraded results due to key point detection errors, synthesis limitations, and global context mismatch. The authors also explored the impact of model size and found that larger models performed worse for face anonymization in the COCO dataset. For full-body anonymization, the standard and multimodal truncation methods improved performance.
To conclude, the study investigated the impact of anonymization on training machine vision models using autonomous vehicle datasets. Face anonymization had minimal effects on instance targeting, while full-body anonymization significantly degraded performance. Realistic anonymization was superior to traditional methods, but not a complete substitute for real data. Privacy protection was highlighted without compromising the performance of the model. The study had limitations in the dependency of annotations and model architectures, which required further research to improve anonymization techniques and address synthesis limitations. The challenges in the synthesis of human figures for anonymization in autonomous vehicles were also highlighted.
review the Paper. Don’t forget to join our 25k+ ML SubReddit, discord channel, and electronic newsletter, where we share the latest AI research news, exciting AI projects, and more. If you have any questions about the article above or if we missed anything, feel free to email us at [email protected]
featured tools Of AI Tools Club
🚀 Check out 100 AI tools at AI Tools Club
Mahmoud is a PhD researcher in machine learning. He also has a
bachelor’s degree in physical sciences and master’s degree in
telecommunication systems and networks. Your current areas of
the research concerns computer vision, stock market prediction and
learning. He produced several scientific articles on the relationship with the person.
identification and study of the robustness and stability of depths
networks
