Year: 2022
Type: dissertation
Abstract: Biometric authentication is becoming widely used, especially in handheld devices.The face is commonly used as a biometric identifier because of its contactless property.Thanks to recent technological advances in both hardware and software, facial biometric identification is well on the way to replacing traditional password authentication, which is inconvenient.Unfortunately, these advanced technologies can also be used to attack biometric authentication systems.For example, deep neural networks can be used to generate realistic images, videos, and speech.Synthetic and manipulated images and videos created in this way are called "deepfakes."They can be used to deceive face authentication systems (by attacking the integrated face recognition system) besides being used to create fake news and to impersonate or harass individuals.As a result, dealing with deepfakes is a vital task in facial biometrics.The countless battles between attackers and defenders have resulted in continuous improvements that make both sides stronger.This philosophy is the principal motivation for the work reported in this thesis.By standing on both the attacker and defender sides, we can simultaneously identify crucial problems in facial biometrics and provide solutions to make it more secure.From the attacker side, we discuss the robustness of face recognition systems under a wolf attack using generated images that could match multiple enrolled user templates.Since being introduced in 2007 for fingerprints, wolf attacks have been widely used against fingerprint-and finger-vein-based authentication systems.Motivated by the use of an evolutionary algorithm along with a generative adversarial network to generate wolf partial fingerprints ("master prints"), we have enhanced the algorithm so that it can generate high-resolution high-quality master faces.Our experimental results demonstrated that, even with limited resources and using only pretrained models available on the Internet, attackers are able to initiate master face attacks.Our generated master faces can be used to attack various types of face recognition systems in white-box, gray-box, and even black-box scenarios.Another contribution is that we identify the limitations of face recognition systems by analyzing the distributions of the face embedding spaces of the systems, then suggest some improvements.From the defender side, we first present a detector that works with various types of computer-generated and manipulated images and videos, commonly known as "deepfakes."The proposed detector uses a capsule network, which is an upgraded version of the traditional convolutional neural network (CNN).The performance of traditional CNNs can be improved by increasing their depth and/or their width, adding more internal connections, or fusing several features or predicted probabilities from multiple CNNs.Consequently, they become bigger, consume more memory and computation iv power, and require more training data.Thanks to the use of a dynamic routing algorithm, our capsule-network-based detector -namely Capsule-Forensics -has fewer parameters than traditional CNNs with similar performance.To further the understanding of the Capsule-Forensics detector, we visualize the activation of its components as a means to improve its explainability.Next, we discuss locating manipulated regions (i.e., performing segmentation), which is important when dealing with fake images and videos.We designed a CNN that uses the multi-task learning approach to simultaneously detect manipulated images and videos and locate the manipulated regions.The information gained by performing one task is shared with the other task, thereby enhancing the performance of all tasks.A semi-supervised learning approach is used to improve the network's generalizability.The network includes an encoder and a Y-shaped decoder.Activation of the encoded features is used for binary classification.The output of one branch of the decoder is used for segmenting the manipulated regions while that of the other branch is used for reconstructing the input, which helps improve overall performance.With this design, fine-tuning the network using only a small amount of data enables it to deal with unseen attacks effectively.
Authors: NGUYEN, Hong Huy, グエン, ホン フイ, Hong Huy, NGUYEN
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7038590311
