Deepfake based backdoor attack against face recognition deep neural networks

• Main Author: Tan, Davis Wen Han
• Other Authors: Chang Chip Hong
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/167031
• Institution: Nanyang Technological University

The recent development and expansion of the field of artificial intelligence has led to a significant increase in the market value of utilizing cutting-edge AI. One branch of AI that has gained widespread attention is Deep Neural Networks (DNNs), known for their outstanding performance in classification tasks. As a result, DNNs have migrated into a variety of application domains. Backdoor attacks have grown to be a significant danger to deep learning models in recent years. A victim model can have a hidden backdoor created that allows a target misclassification to be activated on any poisoned input while still maintaining the classification accuracy on benign inputs by incorporating a small percentage of poisoned samples into the training dataset. Numerous stealthy backdoor generation algorithms have been proposed to make the backdoor trigger unnoticeable to keep it hidden from human observers. Deepfake, makes it very simple for anyone to create highly realistic but fake images, videos, and even voices without the need for advanced technical knowledge. Although this technology was initially designed for use in digital entertainment, the ability to alter content and create new content poses a threat when it is employed for nefarious and malicious purposes, such as disseminating false information. One potential attack is to use Deepfake to embed covert backdoors in DNNs for face recognition. The victim DNNs can correctly identify images of benign faces, but they misclassify targets when given Deepfake's poisoned test images.