Impact of Backdoor Attacks on Face Classification Models through Training Data Poisoning

In this work, we highlight the critical importance of data integrity in training machine learning (ML) models and emphasizes the potential risks associated with outsourcing ML development, data collection, or labeling. Our study investigates the vulnerability of face classification models to backdoor attacks through training data poisoning. We demonstrate that even with a very low presence of poisoned data, it is possible to infect ML models while maintaining high accuracy on normal testing data. The experiments show that backdoor attacks can be as subtle as adding a small occlusion, like a bandage on a face, making it difficult for security personnel to detect in physical spaces. The findings underscore the importance of employing robust validation techniques to maintain reliable and secure models, particularly when using pre-trained models or outsourced data services.